1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * INET An implementation of the TCP/IP protocol suite for the LINUX 4 * operating system. INET is implemented using the BSD Socket 5 * interface as the means of communication with the user level. 6 * 7 * Implementation of the Transmission Control Protocol(TCP). 8 * 9 * Authors: Ross Biro 10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 11 * Mark Evans, <evansmp@uhura.aston.ac.uk> 12 * Corey Minyard <wf-rch!minyard@relay.EU.net> 13 * Florian La Roche, <flla@stud.uni-sb.de> 14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 15 * Linus Torvalds, <torvalds@cs.helsinki.fi> 16 * Alan Cox, <gw4pts@gw4pts.ampr.org> 17 * Matthew Dillon, <dillon@apollo.west.oic.com> 18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 19 * Jorge Cwik, <jorge@laser.satlink.net> 20 * 21 * Fixes: 22 * Alan Cox : Numerous verify_area() calls 23 * Alan Cox : Set the ACK bit on a reset 24 * Alan Cox : Stopped it crashing if it closed while 25 * sk->inuse=1 and was trying to connect 26 * (tcp_err()). 27 * Alan Cox : All icmp error handling was broken 28 * pointers passed where wrong and the 29 * socket was looked up backwards. Nobody 30 * tested any icmp error code obviously. 31 * Alan Cox : tcp_err() now handled properly. It 32 * wakes people on errors. poll 33 * behaves and the icmp error race 34 * has gone by moving it into sock.c 35 * Alan Cox : tcp_send_reset() fixed to work for 36 * everything not just packets for 37 * unknown sockets. 38 * Alan Cox : tcp option processing. 39 * Alan Cox : Reset tweaked (still not 100%) [Had 40 * syn rule wrong] 41 * Herp Rosmanith : More reset fixes 42 * Alan Cox : No longer acks invalid rst frames. 43 * Acking any kind of RST is right out. 44 * Alan Cox : Sets an ignore me flag on an rst 45 * receive otherwise odd bits of prattle 46 * escape still 47 * Alan Cox : Fixed another acking RST frame bug. 48 * Should stop LAN workplace lockups. 49 * Alan Cox : Some tidyups using the new skb list 50 * facilities 51 * Alan Cox : sk->keepopen now seems to work 52 * Alan Cox : Pulls options out correctly on accepts 53 * Alan Cox : Fixed assorted sk->rqueue->next errors 54 * Alan Cox : PSH doesn't end a TCP read. Switched a 55 * bit to skb ops. 56 * Alan Cox : Tidied tcp_data to avoid a potential 57 * nasty. 58 * Alan Cox : Added some better commenting, as the 59 * tcp is hard to follow 60 * Alan Cox : Removed incorrect check for 20 * psh 61 * Michael O'Reilly : ack < copied bug fix. 62 * Johannes Stille : Misc tcp fixes (not all in yet). 63 * Alan Cox : FIN with no memory -> CRASH 64 * Alan Cox : Added socket option proto entries. 65 * Also added awareness of them to accept. 66 * Alan Cox : Added TCP options (SOL_TCP) 67 * Alan Cox : Switched wakeup calls to callbacks, 68 * so the kernel can layer network 69 * sockets. 70 * Alan Cox : Use ip_tos/ip_ttl settings. 71 * Alan Cox : Handle FIN (more) properly (we hope). 72 * Alan Cox : RST frames sent on unsynchronised 73 * state ack error. 74 * Alan Cox : Put in missing check for SYN bit. 75 * Alan Cox : Added tcp_select_window() aka NET2E 76 * window non shrink trick. 77 * Alan Cox : Added a couple of small NET2E timer 78 * fixes 79 * Charles Hedrick : TCP fixes 80 * Toomas Tamm : TCP window fixes 81 * Alan Cox : Small URG fix to rlogin ^C ack fight 82 * Charles Hedrick : Rewrote most of it to actually work 83 * Linus : Rewrote tcp_read() and URG handling 84 * completely 85 * Gerhard Koerting: Fixed some missing timer handling 86 * Matthew Dillon : Reworked TCP machine states as per RFC 87 * Gerhard Koerting: PC/TCP workarounds 88 * Adam Caldwell : Assorted timer/timing errors 89 * Matthew Dillon : Fixed another RST bug 90 * Alan Cox : Move to kernel side addressing changes. 91 * Alan Cox : Beginning work on TCP fastpathing 92 * (not yet usable) 93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine. 94 * Alan Cox : TCP fast path debugging 95 * Alan Cox : Window clamping 96 * Michael Riepe : Bug in tcp_check() 97 * Matt Dillon : More TCP improvements and RST bug fixes 98 * Matt Dillon : Yet more small nasties remove from the 99 * TCP code (Be very nice to this man if 100 * tcp finally works 100%) 8) 101 * Alan Cox : BSD accept semantics. 102 * Alan Cox : Reset on closedown bug. 103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto(). 104 * Michael Pall : Handle poll() after URG properly in 105 * all cases. 106 * Michael Pall : Undo the last fix in tcp_read_urg() 107 * (multi URG PUSH broke rlogin). 108 * Michael Pall : Fix the multi URG PUSH problem in 109 * tcp_readable(), poll() after URG 110 * works now. 111 * Michael Pall : recv(...,MSG_OOB) never blocks in the 112 * BSD api. 113 * Alan Cox : Changed the semantics of sk->socket to 114 * fix a race and a signal problem with 115 * accept() and async I/O. 116 * Alan Cox : Relaxed the rules on tcp_sendto(). 117 * Yury Shevchuk : Really fixed accept() blocking problem. 118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for 119 * clients/servers which listen in on 120 * fixed ports. 121 * Alan Cox : Cleaned the above up and shrank it to 122 * a sensible code size. 123 * Alan Cox : Self connect lockup fix. 124 * Alan Cox : No connect to multicast. 125 * Ross Biro : Close unaccepted children on master 126 * socket close. 127 * Alan Cox : Reset tracing code. 128 * Alan Cox : Spurious resets on shutdown. 129 * Alan Cox : Giant 15 minute/60 second timer error 130 * Alan Cox : Small whoops in polling before an 131 * accept. 132 * Alan Cox : Kept the state trace facility since 133 * it's handy for debugging. 134 * Alan Cox : More reset handler fixes. 135 * Alan Cox : Started rewriting the code based on 136 * the RFC's for other useful protocol 137 * references see: Comer, KA9Q NOS, and 138 * for a reference on the difference 139 * between specifications and how BSD 140 * works see the 4.4lite source. 141 * A.N.Kuznetsov : Don't time wait on completion of tidy 142 * close. 143 * Linus Torvalds : Fin/Shutdown & copied_seq changes. 144 * Linus Torvalds : Fixed BSD port reuse to work first syn 145 * Alan Cox : Reimplemented timers as per the RFC 146 * and using multiple timers for sanity. 147 * Alan Cox : Small bug fixes, and a lot of new 148 * comments. 149 * Alan Cox : Fixed dual reader crash by locking 150 * the buffers (much like datagram.c) 151 * Alan Cox : Fixed stuck sockets in probe. A probe 152 * now gets fed up of retrying without 153 * (even a no space) answer. 154 * Alan Cox : Extracted closing code better 155 * Alan Cox : Fixed the closing state machine to 156 * resemble the RFC. 157 * Alan Cox : More 'per spec' fixes. 158 * Jorge Cwik : Even faster checksumming. 159 * Alan Cox : tcp_data() doesn't ack illegal PSH 160 * only frames. At least one pc tcp stack 161 * generates them. 162 * Alan Cox : Cache last socket. 163 * Alan Cox : Per route irtt. 164 * Matt Day : poll()->select() match BSD precisely on error 165 * Alan Cox : New buffers 166 * Marc Tamsky : Various sk->prot->retransmits and 167 * sk->retransmits misupdating fixed. 168 * Fixed tcp_write_timeout: stuck close, 169 * and TCP syn retries gets used now. 170 * Mark Yarvis : In tcp_read_wakeup(), don't send an 171 * ack if state is TCP_CLOSED. 172 * Alan Cox : Look up device on a retransmit - routes may 173 * change. Doesn't yet cope with MSS shrink right 174 * but it's a start! 175 * Marc Tamsky : Closing in closing fixes. 176 * Mike Shaver : RFC1122 verifications. 177 * Alan Cox : rcv_saddr errors. 178 * Alan Cox : Block double connect(). 179 * Alan Cox : Small hooks for enSKIP. 180 * Alexey Kuznetsov: Path MTU discovery. 181 * Alan Cox : Support soft errors. 182 * Alan Cox : Fix MTU discovery pathological case 183 * when the remote claims no mtu! 184 * Marc Tamsky : TCP_CLOSE fix. 185 * Colin (G3TNE) : Send a reset on syn ack replies in 186 * window but wrong (fixes NT lpd problems) 187 * Pedro Roque : Better TCP window handling, delayed ack. 188 * Joerg Reuter : No modification of locked buffers in 189 * tcp_do_retransmit() 190 * Eric Schenk : Changed receiver side silly window 191 * avoidance algorithm to BSD style 192 * algorithm. This doubles throughput 193 * against machines running Solaris, 194 * and seems to result in general 195 * improvement. 196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD 197 * Willy Konynenberg : Transparent proxying support. 198 * Mike McLagan : Routing by source 199 * Keith Owens : Do proper merging with partial SKB's in 200 * tcp_do_sendmsg to avoid burstiness. 201 * Eric Schenk : Fix fast close down bug with 202 * shutdown() followed by close(). 203 * Andi Kleen : Make poll agree with SIGIO 204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and 205 * lingertime == 0 (RFC 793 ABORT Call) 206 * Hirokazu Takahashi : Use copy_from_user() instead of 207 * csum_and_copy_from_user() if possible. 208 * 209 * Description of States: 210 * 211 * TCP_SYN_SENT sent a connection request, waiting for ack 212 * 213 * TCP_SYN_RECV received a connection request, sent ack, 214 * waiting for final ack in three-way handshake. 215 * 216 * TCP_ESTABLISHED connection established 217 * 218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete 219 * transmission of remaining buffered data 220 * 221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote 222 * to shutdown 223 * 224 * TCP_CLOSING both sides have shutdown but we still have 225 * data we have to finish sending 226 * 227 * TCP_TIME_WAIT timeout to catch resent junk before entering 228 * closed, can only be entered from FIN_WAIT2 229 * or CLOSING. Required because the other end 230 * may not have gotten our last ACK causing it 231 * to retransmit the data packet (which we ignore) 232 * 233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for 234 * us to finish writing our data and to shutdown 235 * (we have to close() to move on to LAST_ACK) 236 * 237 * TCP_LAST_ACK out side has shutdown after remote has 238 * shutdown. There may still be data in our 239 * buffer that we have to finish sending 240 * 241 * TCP_CLOSE socket is finished 242 */ 243 244 #define pr_fmt(fmt) "TCP: " fmt 245 246 #include <crypto/hash.h> 247 #include <linux/kernel.h> 248 #include <linux/module.h> 249 #include <linux/types.h> 250 #include <linux/fcntl.h> 251 #include <linux/poll.h> 252 #include <linux/inet_diag.h> 253 #include <linux/init.h> 254 #include <linux/fs.h> 255 #include <linux/skbuff.h> 256 #include <linux/scatterlist.h> 257 #include <linux/splice.h> 258 #include <linux/net.h> 259 #include <linux/socket.h> 260 #include <linux/random.h> 261 #include <linux/memblock.h> 262 #include <linux/highmem.h> 263 #include <linux/swap.h> 264 #include <linux/cache.h> 265 #include <linux/err.h> 266 #include <linux/time.h> 267 #include <linux/slab.h> 268 #include <linux/errqueue.h> 269 #include <linux/static_key.h> 270 271 #include <net/icmp.h> 272 #include <net/inet_common.h> 273 #include <net/tcp.h> 274 #include <net/mptcp.h> 275 #include <net/xfrm.h> 276 #include <net/ip.h> 277 #include <net/sock.h> 278 279 #include <linux/uaccess.h> 280 #include <asm/ioctls.h> 281 #include <net/busy_poll.h> 282 283 struct percpu_counter tcp_orphan_count; 284 EXPORT_SYMBOL_GPL(tcp_orphan_count); 285 286 long sysctl_tcp_mem[3] __read_mostly; 287 EXPORT_SYMBOL(sysctl_tcp_mem); 288 289 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */ 290 EXPORT_SYMBOL(tcp_memory_allocated); 291 292 #if IS_ENABLED(CONFIG_SMC) 293 DEFINE_STATIC_KEY_FALSE(tcp_have_smc); 294 EXPORT_SYMBOL(tcp_have_smc); 295 #endif 296 297 /* 298 * Current number of TCP sockets. 299 */ 300 struct percpu_counter tcp_sockets_allocated; 301 EXPORT_SYMBOL(tcp_sockets_allocated); 302 303 /* 304 * TCP splice context 305 */ 306 struct tcp_splice_state { 307 struct pipe_inode_info *pipe; 308 size_t len; 309 unsigned int flags; 310 }; 311 312 /* 313 * Pressure flag: try to collapse. 314 * Technical note: it is used by multiple contexts non atomically. 315 * All the __sk_mem_schedule() is of this nature: accounting 316 * is strict, actions are advisory and have some latency. 317 */ 318 unsigned long tcp_memory_pressure __read_mostly; 319 EXPORT_SYMBOL_GPL(tcp_memory_pressure); 320 321 DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key); 322 EXPORT_SYMBOL(tcp_rx_skb_cache_key); 323 324 DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key); 325 326 void tcp_enter_memory_pressure(struct sock *sk) 327 { 328 unsigned long val; 329 330 if (READ_ONCE(tcp_memory_pressure)) 331 return; 332 val = jiffies; 333 334 if (!val) 335 val--; 336 if (!cmpxchg(&tcp_memory_pressure, 0, val)) 337 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 338 } 339 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure); 340 341 void tcp_leave_memory_pressure(struct sock *sk) 342 { 343 unsigned long val; 344 345 if (!READ_ONCE(tcp_memory_pressure)) 346 return; 347 val = xchg(&tcp_memory_pressure, 0); 348 if (val) 349 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO, 350 jiffies_to_msecs(jiffies - val)); 351 } 352 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure); 353 354 /* Convert seconds to retransmits based on initial and max timeout */ 355 static u8 secs_to_retrans(int seconds, int timeout, int rto_max) 356 { 357 u8 res = 0; 358 359 if (seconds > 0) { 360 int period = timeout; 361 362 res = 1; 363 while (seconds > period && res < 255) { 364 res++; 365 timeout <<= 1; 366 if (timeout > rto_max) 367 timeout = rto_max; 368 period += timeout; 369 } 370 } 371 return res; 372 } 373 374 /* Convert retransmits to seconds based on initial and max timeout */ 375 static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 376 { 377 int period = 0; 378 379 if (retrans > 0) { 380 period = timeout; 381 while (--retrans) { 382 timeout <<= 1; 383 if (timeout > rto_max) 384 timeout = rto_max; 385 period += timeout; 386 } 387 } 388 return period; 389 } 390 391 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp) 392 { 393 u32 rate = READ_ONCE(tp->rate_delivered); 394 u32 intv = READ_ONCE(tp->rate_interval_us); 395 u64 rate64 = 0; 396 397 if (rate && intv) { 398 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC; 399 do_div(rate64, intv); 400 } 401 return rate64; 402 } 403 404 /* Address-family independent initialization for a tcp_sock. 405 * 406 * NOTE: A lot of things set to zero explicitly by call to 407 * sk_alloc() so need not be done here. 408 */ 409 void tcp_init_sock(struct sock *sk) 410 { 411 struct inet_connection_sock *icsk = inet_csk(sk); 412 struct tcp_sock *tp = tcp_sk(sk); 413 414 tp->out_of_order_queue = RB_ROOT; 415 sk->tcp_rtx_queue = RB_ROOT; 416 tcp_init_xmit_timers(sk); 417 INIT_LIST_HEAD(&tp->tsq_node); 418 INIT_LIST_HEAD(&tp->tsorted_sent_queue); 419 420 icsk->icsk_rto = TCP_TIMEOUT_INIT; 421 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 422 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U); 423 424 /* So many TCP implementations out there (incorrectly) count the 425 * initial SYN frame in their delayed-ACK and congestion control 426 * algorithms that we must have the following bandaid to talk 427 * efficiently to them. -DaveM 428 */ 429 tp->snd_cwnd = TCP_INIT_CWND; 430 431 /* There's a bubble in the pipe until at least the first ACK. */ 432 tp->app_limited = ~0U; 433 434 /* See draft-stevens-tcpca-spec-01 for discussion of the 435 * initialization of these values. 436 */ 437 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 438 tp->snd_cwnd_clamp = ~0; 439 tp->mss_cache = TCP_MSS_DEFAULT; 440 441 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering; 442 tcp_assign_congestion_control(sk); 443 444 tp->tsoffset = 0; 445 tp->rack.reo_wnd_steps = 1; 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 WRITE_ONCE(sk->sk_sndbuf, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]); 453 WRITE_ONCE(sk->sk_rcvbuf, sock_net(sk)->ipv4.sysctl_tcp_rmem[1]); 454 455 sk_sockets_allocated_inc(sk); 456 sk->sk_route_forced_caps = NETIF_F_GSO; 457 } 458 EXPORT_SYMBOL(tcp_init_sock); 459 460 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags) 461 { 462 struct sk_buff *skb = tcp_write_queue_tail(sk); 463 464 if (tsflags && skb) { 465 struct skb_shared_info *shinfo = skb_shinfo(skb); 466 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 467 468 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags); 469 if (tsflags & SOF_TIMESTAMPING_TX_ACK) 470 tcb->txstamp_ack = 1; 471 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK) 472 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1; 473 } 474 } 475 476 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp, 477 int target, struct sock *sk) 478 { 479 int avail = READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq); 480 481 if (avail > 0) { 482 if (avail >= target) 483 return true; 484 if (tcp_rmem_pressure(sk)) 485 return true; 486 } 487 if (sk->sk_prot->stream_memory_read) 488 return sk->sk_prot->stream_memory_read(sk); 489 return false; 490 } 491 492 /* 493 * Wait for a TCP event. 494 * 495 * Note that we don't need to lock the socket, as the upper poll layers 496 * take care of normal races (between the test and the event) and we don't 497 * go look at any of the socket buffers directly. 498 */ 499 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 500 { 501 __poll_t mask; 502 struct sock *sk = sock->sk; 503 const struct tcp_sock *tp = tcp_sk(sk); 504 int state; 505 506 sock_poll_wait(file, sock, wait); 507 508 state = inet_sk_state_load(sk); 509 if (state == TCP_LISTEN) 510 return inet_csk_listen_poll(sk); 511 512 /* Socket is not locked. We are protected from async events 513 * by poll logic and correct handling of state changes 514 * made by other threads is impossible in any case. 515 */ 516 517 mask = 0; 518 519 /* 520 * EPOLLHUP is certainly not done right. But poll() doesn't 521 * have a notion of HUP in just one direction, and for a 522 * socket the read side is more interesting. 523 * 524 * Some poll() documentation says that EPOLLHUP is incompatible 525 * with the EPOLLOUT/POLLWR flags, so somebody should check this 526 * all. But careful, it tends to be safer to return too many 527 * bits than too few, and you can easily break real applications 528 * if you don't tell them that something has hung up! 529 * 530 * Check-me. 531 * 532 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and 533 * our fs/select.c). It means that after we received EOF, 534 * poll always returns immediately, making impossible poll() on write() 535 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP 536 * if and only if shutdown has been made in both directions. 537 * Actually, it is interesting to look how Solaris and DUX 538 * solve this dilemma. I would prefer, if EPOLLHUP were maskable, 539 * then we could set it on SND_SHUTDOWN. BTW examples given 540 * in Stevens' books assume exactly this behaviour, it explains 541 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK 542 * 543 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 544 * blocking on fresh not-connected or disconnected socket. --ANK 545 */ 546 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 547 mask |= EPOLLHUP; 548 if (sk->sk_shutdown & RCV_SHUTDOWN) 549 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 550 551 /* Connected or passive Fast Open socket? */ 552 if (state != TCP_SYN_SENT && 553 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) { 554 int target = sock_rcvlowat(sk, 0, INT_MAX); 555 556 if (READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) && 557 !sock_flag(sk, SOCK_URGINLINE) && 558 tp->urg_data) 559 target++; 560 561 if (tcp_stream_is_readable(tp, target, sk)) 562 mask |= EPOLLIN | EPOLLRDNORM; 563 564 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) { 565 if (sk_stream_is_writeable(sk)) { 566 mask |= EPOLLOUT | EPOLLWRNORM; 567 } else { /* send SIGIO later */ 568 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk); 569 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 570 571 /* Race breaker. If space is freed after 572 * wspace test but before the flags are set, 573 * IO signal will be lost. Memory barrier 574 * pairs with the input side. 575 */ 576 smp_mb__after_atomic(); 577 if (sk_stream_is_writeable(sk)) 578 mask |= EPOLLOUT | EPOLLWRNORM; 579 } 580 } else 581 mask |= EPOLLOUT | EPOLLWRNORM; 582 583 if (tp->urg_data & TCP_URG_VALID) 584 mask |= EPOLLPRI; 585 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) { 586 /* Active TCP fastopen socket with defer_connect 587 * Return EPOLLOUT so application can call write() 588 * in order for kernel to generate SYN+data 589 */ 590 mask |= EPOLLOUT | EPOLLWRNORM; 591 } 592 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 593 smp_rmb(); 594 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue)) 595 mask |= EPOLLERR; 596 597 return mask; 598 } 599 EXPORT_SYMBOL(tcp_poll); 600 601 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 602 { 603 struct tcp_sock *tp = tcp_sk(sk); 604 int answ; 605 bool slow; 606 607 switch (cmd) { 608 case SIOCINQ: 609 if (sk->sk_state == TCP_LISTEN) 610 return -EINVAL; 611 612 slow = lock_sock_fast(sk); 613 answ = tcp_inq(sk); 614 unlock_sock_fast(sk, slow); 615 break; 616 case SIOCATMARK: 617 answ = tp->urg_data && 618 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq); 619 break; 620 case SIOCOUTQ: 621 if (sk->sk_state == TCP_LISTEN) 622 return -EINVAL; 623 624 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 625 answ = 0; 626 else 627 answ = READ_ONCE(tp->write_seq) - tp->snd_una; 628 break; 629 case SIOCOUTQNSD: 630 if (sk->sk_state == TCP_LISTEN) 631 return -EINVAL; 632 633 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 634 answ = 0; 635 else 636 answ = READ_ONCE(tp->write_seq) - 637 READ_ONCE(tp->snd_nxt); 638 break; 639 default: 640 return -ENOIOCTLCMD; 641 } 642 643 return put_user(answ, (int __user *)arg); 644 } 645 EXPORT_SYMBOL(tcp_ioctl); 646 647 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 648 { 649 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 650 tp->pushed_seq = tp->write_seq; 651 } 652 653 static inline bool forced_push(const struct tcp_sock *tp) 654 { 655 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 656 } 657 658 static void skb_entail(struct sock *sk, struct sk_buff *skb) 659 { 660 struct tcp_sock *tp = tcp_sk(sk); 661 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 662 663 skb->csum = 0; 664 tcb->seq = tcb->end_seq = tp->write_seq; 665 tcb->tcp_flags = TCPHDR_ACK; 666 tcb->sacked = 0; 667 __skb_header_release(skb); 668 tcp_add_write_queue_tail(sk, skb); 669 sk_wmem_queued_add(sk, skb->truesize); 670 sk_mem_charge(sk, skb->truesize); 671 if (tp->nonagle & TCP_NAGLE_PUSH) 672 tp->nonagle &= ~TCP_NAGLE_PUSH; 673 674 tcp_slow_start_after_idle_check(sk); 675 } 676 677 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 678 { 679 if (flags & MSG_OOB) 680 tp->snd_up = tp->write_seq; 681 } 682 683 /* If a not yet filled skb is pushed, do not send it if 684 * we have data packets in Qdisc or NIC queues : 685 * Because TX completion will happen shortly, it gives a chance 686 * to coalesce future sendmsg() payload into this skb, without 687 * need for a timer, and with no latency trade off. 688 * As packets containing data payload have a bigger truesize 689 * than pure acks (dataless) packets, the last checks prevent 690 * autocorking if we only have an ACK in Qdisc/NIC queues, 691 * or if TX completion was delayed after we processed ACK packet. 692 */ 693 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb, 694 int size_goal) 695 { 696 return skb->len < size_goal && 697 sock_net(sk)->ipv4.sysctl_tcp_autocorking && 698 !tcp_rtx_queue_empty(sk) && 699 refcount_read(&sk->sk_wmem_alloc) > skb->truesize; 700 } 701 702 void tcp_push(struct sock *sk, int flags, int mss_now, 703 int nonagle, int size_goal) 704 { 705 struct tcp_sock *tp = tcp_sk(sk); 706 struct sk_buff *skb; 707 708 skb = tcp_write_queue_tail(sk); 709 if (!skb) 710 return; 711 if (!(flags & MSG_MORE) || forced_push(tp)) 712 tcp_mark_push(tp, skb); 713 714 tcp_mark_urg(tp, flags); 715 716 if (tcp_should_autocork(sk, skb, size_goal)) { 717 718 /* avoid atomic op if TSQ_THROTTLED bit is already set */ 719 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) { 720 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING); 721 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 722 } 723 /* It is possible TX completion already happened 724 * before we set TSQ_THROTTLED. 725 */ 726 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize) 727 return; 728 } 729 730 if (flags & MSG_MORE) 731 nonagle = TCP_NAGLE_CORK; 732 733 __tcp_push_pending_frames(sk, mss_now, nonagle); 734 } 735 736 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 737 unsigned int offset, size_t len) 738 { 739 struct tcp_splice_state *tss = rd_desc->arg.data; 740 int ret; 741 742 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe, 743 min(rd_desc->count, len), tss->flags); 744 if (ret > 0) 745 rd_desc->count -= ret; 746 return ret; 747 } 748 749 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 750 { 751 /* Store TCP splice context information in read_descriptor_t. */ 752 read_descriptor_t rd_desc = { 753 .arg.data = tss, 754 .count = tss->len, 755 }; 756 757 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 758 } 759 760 /** 761 * tcp_splice_read - splice data from TCP socket to a pipe 762 * @sock: socket to splice from 763 * @ppos: position (not valid) 764 * @pipe: pipe to splice to 765 * @len: number of bytes to splice 766 * @flags: splice modifier flags 767 * 768 * Description: 769 * Will read pages from given socket and fill them into a pipe. 770 * 771 **/ 772 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 773 struct pipe_inode_info *pipe, size_t len, 774 unsigned int flags) 775 { 776 struct sock *sk = sock->sk; 777 struct tcp_splice_state tss = { 778 .pipe = pipe, 779 .len = len, 780 .flags = flags, 781 }; 782 long timeo; 783 ssize_t spliced; 784 int ret; 785 786 sock_rps_record_flow(sk); 787 /* 788 * We can't seek on a socket input 789 */ 790 if (unlikely(*ppos)) 791 return -ESPIPE; 792 793 ret = spliced = 0; 794 795 lock_sock(sk); 796 797 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 798 while (tss.len) { 799 ret = __tcp_splice_read(sk, &tss); 800 if (ret < 0) 801 break; 802 else if (!ret) { 803 if (spliced) 804 break; 805 if (sock_flag(sk, SOCK_DONE)) 806 break; 807 if (sk->sk_err) { 808 ret = sock_error(sk); 809 break; 810 } 811 if (sk->sk_shutdown & RCV_SHUTDOWN) 812 break; 813 if (sk->sk_state == TCP_CLOSE) { 814 /* 815 * This occurs when user tries to read 816 * from never connected socket. 817 */ 818 ret = -ENOTCONN; 819 break; 820 } 821 if (!timeo) { 822 ret = -EAGAIN; 823 break; 824 } 825 /* if __tcp_splice_read() got nothing while we have 826 * an skb in receive queue, we do not want to loop. 827 * This might happen with URG data. 828 */ 829 if (!skb_queue_empty(&sk->sk_receive_queue)) 830 break; 831 sk_wait_data(sk, &timeo, NULL); 832 if (signal_pending(current)) { 833 ret = sock_intr_errno(timeo); 834 break; 835 } 836 continue; 837 } 838 tss.len -= ret; 839 spliced += ret; 840 841 if (!timeo) 842 break; 843 release_sock(sk); 844 lock_sock(sk); 845 846 if (sk->sk_err || sk->sk_state == TCP_CLOSE || 847 (sk->sk_shutdown & RCV_SHUTDOWN) || 848 signal_pending(current)) 849 break; 850 } 851 852 release_sock(sk); 853 854 if (spliced) 855 return spliced; 856 857 return ret; 858 } 859 EXPORT_SYMBOL(tcp_splice_read); 860 861 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp, 862 bool force_schedule) 863 { 864 struct sk_buff *skb; 865 866 if (likely(!size)) { 867 skb = sk->sk_tx_skb_cache; 868 if (skb) { 869 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb)); 870 sk->sk_tx_skb_cache = NULL; 871 pskb_trim(skb, 0); 872 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 873 skb_shinfo(skb)->tx_flags = 0; 874 memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb)); 875 return skb; 876 } 877 } 878 /* The TCP header must be at least 32-bit aligned. */ 879 size = ALIGN(size, 4); 880 881 if (unlikely(tcp_under_memory_pressure(sk))) 882 sk_mem_reclaim_partial(sk); 883 884 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp); 885 if (likely(skb)) { 886 bool mem_scheduled; 887 888 if (force_schedule) { 889 mem_scheduled = true; 890 sk_forced_mem_schedule(sk, skb->truesize); 891 } else { 892 mem_scheduled = sk_wmem_schedule(sk, skb->truesize); 893 } 894 if (likely(mem_scheduled)) { 895 skb_reserve(skb, sk->sk_prot->max_header); 896 /* 897 * Make sure that we have exactly size bytes 898 * available to the caller, no more, no less. 899 */ 900 skb->reserved_tailroom = skb->end - skb->tail - size; 901 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 902 return skb; 903 } 904 __kfree_skb(skb); 905 } else { 906 sk->sk_prot->enter_memory_pressure(sk); 907 sk_stream_moderate_sndbuf(sk); 908 } 909 return NULL; 910 } 911 912 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 913 int large_allowed) 914 { 915 struct tcp_sock *tp = tcp_sk(sk); 916 u32 new_size_goal, size_goal; 917 918 if (!large_allowed) 919 return mss_now; 920 921 /* Note : tcp_tso_autosize() will eventually split this later */ 922 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER; 923 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal); 924 925 /* We try hard to avoid divides here */ 926 size_goal = tp->gso_segs * mss_now; 927 if (unlikely(new_size_goal < size_goal || 928 new_size_goal >= size_goal + mss_now)) { 929 tp->gso_segs = min_t(u16, new_size_goal / mss_now, 930 sk->sk_gso_max_segs); 931 size_goal = tp->gso_segs * mss_now; 932 } 933 934 return max(size_goal, mss_now); 935 } 936 937 int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 938 { 939 int mss_now; 940 941 mss_now = tcp_current_mss(sk); 942 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 943 944 return mss_now; 945 } 946 947 /* In some cases, both sendpage() and sendmsg() could have added 948 * an skb to the write queue, but failed adding payload on it. 949 * We need to remove it to consume less memory, but more 950 * importantly be able to generate EPOLLOUT for Edge Trigger epoll() 951 * users. 952 */ 953 static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb) 954 { 955 if (skb && !skb->len) { 956 tcp_unlink_write_queue(skb, sk); 957 if (tcp_write_queue_empty(sk)) 958 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 959 sk_wmem_free_skb(sk, skb); 960 } 961 } 962 963 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset, 964 size_t size, int flags) 965 { 966 struct tcp_sock *tp = tcp_sk(sk); 967 int mss_now, size_goal; 968 int err; 969 ssize_t copied; 970 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 971 972 if (IS_ENABLED(CONFIG_DEBUG_VM) && 973 WARN_ONCE(PageSlab(page), "page must not be a Slab one")) 974 return -EINVAL; 975 976 /* Wait for a connection to finish. One exception is TCP Fast Open 977 * (passive side) where data is allowed to be sent before a connection 978 * is fully established. 979 */ 980 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 981 !tcp_passive_fastopen(sk)) { 982 err = sk_stream_wait_connect(sk, &timeo); 983 if (err != 0) 984 goto out_err; 985 } 986 987 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 988 989 mss_now = tcp_send_mss(sk, &size_goal, flags); 990 copied = 0; 991 992 err = -EPIPE; 993 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 994 goto out_err; 995 996 while (size > 0) { 997 struct sk_buff *skb = tcp_write_queue_tail(sk); 998 int copy, i; 999 bool can_coalesce; 1000 1001 if (!skb || (copy = size_goal - skb->len) <= 0 || 1002 !tcp_skb_can_collapse_to(skb)) { 1003 new_segment: 1004 if (!sk_stream_memory_free(sk)) 1005 goto wait_for_sndbuf; 1006 1007 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, 1008 tcp_rtx_and_write_queues_empty(sk)); 1009 if (!skb) 1010 goto wait_for_memory; 1011 1012 #ifdef CONFIG_TLS_DEVICE 1013 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED); 1014 #endif 1015 skb_entail(sk, skb); 1016 copy = size_goal; 1017 } 1018 1019 if (copy > size) 1020 copy = size; 1021 1022 i = skb_shinfo(skb)->nr_frags; 1023 can_coalesce = skb_can_coalesce(skb, i, page, offset); 1024 if (!can_coalesce && i >= sysctl_max_skb_frags) { 1025 tcp_mark_push(tp, skb); 1026 goto new_segment; 1027 } 1028 if (!sk_wmem_schedule(sk, copy)) 1029 goto wait_for_memory; 1030 1031 if (can_coalesce) { 1032 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1033 } else { 1034 get_page(page); 1035 skb_fill_page_desc(skb, i, page, offset, copy); 1036 } 1037 1038 if (!(flags & MSG_NO_SHARED_FRAGS)) 1039 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG; 1040 1041 skb->len += copy; 1042 skb->data_len += copy; 1043 skb->truesize += copy; 1044 sk_wmem_queued_add(sk, copy); 1045 sk_mem_charge(sk, copy); 1046 skb->ip_summed = CHECKSUM_PARTIAL; 1047 WRITE_ONCE(tp->write_seq, tp->write_seq + copy); 1048 TCP_SKB_CB(skb)->end_seq += copy; 1049 tcp_skb_pcount_set(skb, 0); 1050 1051 if (!copied) 1052 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1053 1054 copied += copy; 1055 offset += copy; 1056 size -= copy; 1057 if (!size) 1058 goto out; 1059 1060 if (skb->len < size_goal || (flags & MSG_OOB)) 1061 continue; 1062 1063 if (forced_push(tp)) { 1064 tcp_mark_push(tp, skb); 1065 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1066 } else if (skb == tcp_send_head(sk)) 1067 tcp_push_one(sk, mss_now); 1068 continue; 1069 1070 wait_for_sndbuf: 1071 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1072 wait_for_memory: 1073 tcp_push(sk, flags & ~MSG_MORE, mss_now, 1074 TCP_NAGLE_PUSH, size_goal); 1075 1076 err = sk_stream_wait_memory(sk, &timeo); 1077 if (err != 0) 1078 goto do_error; 1079 1080 mss_now = tcp_send_mss(sk, &size_goal, flags); 1081 } 1082 1083 out: 1084 if (copied) { 1085 tcp_tx_timestamp(sk, sk->sk_tsflags); 1086 if (!(flags & MSG_SENDPAGE_NOTLAST)) 1087 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 1088 } 1089 return copied; 1090 1091 do_error: 1092 tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk)); 1093 if (copied) 1094 goto out; 1095 out_err: 1096 /* make sure we wake any epoll edge trigger waiter */ 1097 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) { 1098 sk->sk_write_space(sk); 1099 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 1100 } 1101 return sk_stream_error(sk, flags, err); 1102 } 1103 EXPORT_SYMBOL_GPL(do_tcp_sendpages); 1104 1105 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset, 1106 size_t size, int flags) 1107 { 1108 if (!(sk->sk_route_caps & NETIF_F_SG)) 1109 return sock_no_sendpage_locked(sk, page, offset, size, flags); 1110 1111 tcp_rate_check_app_limited(sk); /* is sending application-limited? */ 1112 1113 return do_tcp_sendpages(sk, page, offset, size, flags); 1114 } 1115 EXPORT_SYMBOL_GPL(tcp_sendpage_locked); 1116 1117 int tcp_sendpage(struct sock *sk, struct page *page, int offset, 1118 size_t size, int flags) 1119 { 1120 int ret; 1121 1122 lock_sock(sk); 1123 ret = tcp_sendpage_locked(sk, page, offset, size, flags); 1124 release_sock(sk); 1125 1126 return ret; 1127 } 1128 EXPORT_SYMBOL(tcp_sendpage); 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 struct ubuf_info *uarg) 1141 { 1142 struct tcp_sock *tp = tcp_sk(sk); 1143 struct inet_sock *inet = inet_sk(sk); 1144 struct sockaddr *uaddr = msg->msg_name; 1145 int err, flags; 1146 1147 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) || 1148 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) && 1149 uaddr->sa_family == AF_UNSPEC)) 1150 return -EOPNOTSUPP; 1151 if (tp->fastopen_req) 1152 return -EALREADY; /* Another Fast Open is in progress */ 1153 1154 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request), 1155 sk->sk_allocation); 1156 if (unlikely(!tp->fastopen_req)) 1157 return -ENOBUFS; 1158 tp->fastopen_req->data = msg; 1159 tp->fastopen_req->size = size; 1160 tp->fastopen_req->uarg = uarg; 1161 1162 if (inet->defer_connect) { 1163 err = tcp_connect(sk); 1164 /* Same failure procedure as in tcp_v4/6_connect */ 1165 if (err) { 1166 tcp_set_state(sk, TCP_CLOSE); 1167 inet->inet_dport = 0; 1168 sk->sk_route_caps = 0; 1169 } 1170 } 1171 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0; 1172 err = __inet_stream_connect(sk->sk_socket, uaddr, 1173 msg->msg_namelen, flags, 1); 1174 /* fastopen_req could already be freed in __inet_stream_connect 1175 * if the connection times out or gets rst 1176 */ 1177 if (tp->fastopen_req) { 1178 *copied = tp->fastopen_req->copied; 1179 tcp_free_fastopen_req(tp); 1180 inet->defer_connect = 0; 1181 } 1182 return err; 1183 } 1184 1185 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size) 1186 { 1187 struct tcp_sock *tp = tcp_sk(sk); 1188 struct ubuf_info *uarg = NULL; 1189 struct sk_buff *skb; 1190 struct sockcm_cookie sockc; 1191 int flags, err, copied = 0; 1192 int mss_now = 0, size_goal, copied_syn = 0; 1193 int process_backlog = 0; 1194 bool zc = false; 1195 long timeo; 1196 1197 flags = msg->msg_flags; 1198 1199 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) { 1200 skb = tcp_write_queue_tail(sk); 1201 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb)); 1202 if (!uarg) { 1203 err = -ENOBUFS; 1204 goto out_err; 1205 } 1206 1207 zc = sk->sk_route_caps & NETIF_F_SG; 1208 if (!zc) 1209 uarg->zerocopy = 0; 1210 } 1211 1212 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) && 1213 !tp->repair) { 1214 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg); 1215 if (err == -EINPROGRESS && copied_syn > 0) 1216 goto out; 1217 else if (err) 1218 goto out_err; 1219 } 1220 1221 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 1222 1223 tcp_rate_check_app_limited(sk); /* is sending application-limited? */ 1224 1225 /* Wait for a connection to finish. One exception is TCP Fast Open 1226 * (passive side) where data is allowed to be sent before a connection 1227 * is fully established. 1228 */ 1229 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 1230 !tcp_passive_fastopen(sk)) { 1231 err = sk_stream_wait_connect(sk, &timeo); 1232 if (err != 0) 1233 goto do_error; 1234 } 1235 1236 if (unlikely(tp->repair)) { 1237 if (tp->repair_queue == TCP_RECV_QUEUE) { 1238 copied = tcp_send_rcvq(sk, msg, size); 1239 goto out_nopush; 1240 } 1241 1242 err = -EINVAL; 1243 if (tp->repair_queue == TCP_NO_QUEUE) 1244 goto out_err; 1245 1246 /* 'common' sending to sendq */ 1247 } 1248 1249 sockcm_init(&sockc, sk); 1250 if (msg->msg_controllen) { 1251 err = sock_cmsg_send(sk, msg, &sockc); 1252 if (unlikely(err)) { 1253 err = -EINVAL; 1254 goto out_err; 1255 } 1256 } 1257 1258 /* This should be in poll */ 1259 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 1260 1261 /* Ok commence sending. */ 1262 copied = 0; 1263 1264 restart: 1265 mss_now = tcp_send_mss(sk, &size_goal, flags); 1266 1267 err = -EPIPE; 1268 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 1269 goto do_error; 1270 1271 while (msg_data_left(msg)) { 1272 int copy = 0; 1273 1274 skb = tcp_write_queue_tail(sk); 1275 if (skb) 1276 copy = size_goal - skb->len; 1277 1278 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) { 1279 bool first_skb; 1280 1281 new_segment: 1282 if (!sk_stream_memory_free(sk)) 1283 goto wait_for_sndbuf; 1284 1285 if (unlikely(process_backlog >= 16)) { 1286 process_backlog = 0; 1287 if (sk_flush_backlog(sk)) 1288 goto restart; 1289 } 1290 first_skb = tcp_rtx_and_write_queues_empty(sk); 1291 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, 1292 first_skb); 1293 if (!skb) 1294 goto wait_for_memory; 1295 1296 process_backlog++; 1297 skb->ip_summed = CHECKSUM_PARTIAL; 1298 1299 skb_entail(sk, skb); 1300 copy = size_goal; 1301 1302 /* All packets are restored as if they have 1303 * already been sent. skb_mstamp_ns isn't set to 1304 * avoid wrong rtt estimation. 1305 */ 1306 if (tp->repair) 1307 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED; 1308 } 1309 1310 /* Try to append data to the end of skb. */ 1311 if (copy > msg_data_left(msg)) 1312 copy = msg_data_left(msg); 1313 1314 /* Where to copy to? */ 1315 if (skb_availroom(skb) > 0 && !zc) { 1316 /* We have some space in skb head. Superb! */ 1317 copy = min_t(int, copy, skb_availroom(skb)); 1318 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy); 1319 if (err) 1320 goto do_fault; 1321 } else if (!zc) { 1322 bool merge = true; 1323 int i = skb_shinfo(skb)->nr_frags; 1324 struct page_frag *pfrag = sk_page_frag(sk); 1325 1326 if (!sk_page_frag_refill(sk, pfrag)) 1327 goto wait_for_memory; 1328 1329 if (!skb_can_coalesce(skb, i, pfrag->page, 1330 pfrag->offset)) { 1331 if (i >= sysctl_max_skb_frags) { 1332 tcp_mark_push(tp, skb); 1333 goto new_segment; 1334 } 1335 merge = false; 1336 } 1337 1338 copy = min_t(int, copy, pfrag->size - pfrag->offset); 1339 1340 if (!sk_wmem_schedule(sk, copy)) 1341 goto wait_for_memory; 1342 1343 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb, 1344 pfrag->page, 1345 pfrag->offset, 1346 copy); 1347 if (err) 1348 goto do_error; 1349 1350 /* Update the skb. */ 1351 if (merge) { 1352 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1353 } else { 1354 skb_fill_page_desc(skb, i, pfrag->page, 1355 pfrag->offset, copy); 1356 page_ref_inc(pfrag->page); 1357 } 1358 pfrag->offset += copy; 1359 } else { 1360 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg); 1361 if (err == -EMSGSIZE || err == -EEXIST) { 1362 tcp_mark_push(tp, skb); 1363 goto new_segment; 1364 } 1365 if (err < 0) 1366 goto do_error; 1367 copy = err; 1368 } 1369 1370 if (!copied) 1371 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1372 1373 WRITE_ONCE(tp->write_seq, tp->write_seq + copy); 1374 TCP_SKB_CB(skb)->end_seq += copy; 1375 tcp_skb_pcount_set(skb, 0); 1376 1377 copied += copy; 1378 if (!msg_data_left(msg)) { 1379 if (unlikely(flags & MSG_EOR)) 1380 TCP_SKB_CB(skb)->eor = 1; 1381 goto out; 1382 } 1383 1384 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair)) 1385 continue; 1386 1387 if (forced_push(tp)) { 1388 tcp_mark_push(tp, skb); 1389 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1390 } else if (skb == tcp_send_head(sk)) 1391 tcp_push_one(sk, mss_now); 1392 continue; 1393 1394 wait_for_sndbuf: 1395 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1396 wait_for_memory: 1397 if (copied) 1398 tcp_push(sk, flags & ~MSG_MORE, mss_now, 1399 TCP_NAGLE_PUSH, size_goal); 1400 1401 err = sk_stream_wait_memory(sk, &timeo); 1402 if (err != 0) 1403 goto do_error; 1404 1405 mss_now = tcp_send_mss(sk, &size_goal, flags); 1406 } 1407 1408 out: 1409 if (copied) { 1410 tcp_tx_timestamp(sk, sockc.tsflags); 1411 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 1412 } 1413 out_nopush: 1414 sock_zerocopy_put(uarg); 1415 return copied + copied_syn; 1416 1417 do_error: 1418 skb = tcp_write_queue_tail(sk); 1419 do_fault: 1420 tcp_remove_empty_skb(sk, skb); 1421 1422 if (copied + copied_syn) 1423 goto out; 1424 out_err: 1425 sock_zerocopy_put_abort(uarg, true); 1426 err = sk_stream_error(sk, flags, err); 1427 /* make sure we wake any epoll edge trigger waiter */ 1428 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) { 1429 sk->sk_write_space(sk); 1430 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 1431 } 1432 return err; 1433 } 1434 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked); 1435 1436 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 1437 { 1438 int ret; 1439 1440 lock_sock(sk); 1441 ret = tcp_sendmsg_locked(sk, msg, size); 1442 release_sock(sk); 1443 1444 return ret; 1445 } 1446 EXPORT_SYMBOL(tcp_sendmsg); 1447 1448 /* 1449 * Handle reading urgent data. BSD has very simple semantics for 1450 * this, no blocking and very strange errors 8) 1451 */ 1452 1453 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1454 { 1455 struct tcp_sock *tp = tcp_sk(sk); 1456 1457 /* No URG data to read. */ 1458 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1459 tp->urg_data == TCP_URG_READ) 1460 return -EINVAL; /* Yes this is right ! */ 1461 1462 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1463 return -ENOTCONN; 1464 1465 if (tp->urg_data & TCP_URG_VALID) { 1466 int err = 0; 1467 char c = tp->urg_data; 1468 1469 if (!(flags & MSG_PEEK)) 1470 tp->urg_data = TCP_URG_READ; 1471 1472 /* Read urgent data. */ 1473 msg->msg_flags |= MSG_OOB; 1474 1475 if (len > 0) { 1476 if (!(flags & MSG_TRUNC)) 1477 err = memcpy_to_msg(msg, &c, 1); 1478 len = 1; 1479 } else 1480 msg->msg_flags |= MSG_TRUNC; 1481 1482 return err ? -EFAULT : len; 1483 } 1484 1485 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1486 return 0; 1487 1488 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1489 * the available implementations agree in this case: 1490 * this call should never block, independent of the 1491 * blocking state of the socket. 1492 * Mike <pall@rz.uni-karlsruhe.de> 1493 */ 1494 return -EAGAIN; 1495 } 1496 1497 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len) 1498 { 1499 struct sk_buff *skb; 1500 int copied = 0, err = 0; 1501 1502 /* XXX -- need to support SO_PEEK_OFF */ 1503 1504 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 1505 err = skb_copy_datagram_msg(skb, 0, msg, skb->len); 1506 if (err) 1507 return err; 1508 copied += skb->len; 1509 } 1510 1511 skb_queue_walk(&sk->sk_write_queue, skb) { 1512 err = skb_copy_datagram_msg(skb, 0, msg, skb->len); 1513 if (err) 1514 break; 1515 1516 copied += skb->len; 1517 } 1518 1519 return err ?: copied; 1520 } 1521 1522 /* Clean up the receive buffer for full frames taken by the user, 1523 * then send an ACK if necessary. COPIED is the number of bytes 1524 * tcp_recvmsg has given to the user so far, it speeds up the 1525 * calculation of whether or not we must ACK for the sake of 1526 * a window update. 1527 */ 1528 static void tcp_cleanup_rbuf(struct sock *sk, int copied) 1529 { 1530 struct tcp_sock *tp = tcp_sk(sk); 1531 bool time_to_ack = false; 1532 1533 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1534 1535 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1536 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1537 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1538 1539 if (inet_csk_ack_scheduled(sk)) { 1540 const struct inet_connection_sock *icsk = inet_csk(sk); 1541 /* Delayed ACKs frequently hit locked sockets during bulk 1542 * receive. */ 1543 if (icsk->icsk_ack.blocked || 1544 /* Once-per-two-segments ACK was not sent by tcp_input.c */ 1545 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1546 /* 1547 * If this read emptied read buffer, we send ACK, if 1548 * connection is not bidirectional, user drained 1549 * receive buffer and there was a small segment 1550 * in queue. 1551 */ 1552 (copied > 0 && 1553 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1554 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1555 !inet_csk_in_pingpong_mode(sk))) && 1556 !atomic_read(&sk->sk_rmem_alloc))) 1557 time_to_ack = true; 1558 } 1559 1560 /* We send an ACK if we can now advertise a non-zero window 1561 * which has been raised "significantly". 1562 * 1563 * Even if window raised up to infinity, do not send window open ACK 1564 * in states, where we will not receive more. It is useless. 1565 */ 1566 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1567 __u32 rcv_window_now = tcp_receive_window(tp); 1568 1569 /* Optimize, __tcp_select_window() is not cheap. */ 1570 if (2*rcv_window_now <= tp->window_clamp) { 1571 __u32 new_window = __tcp_select_window(sk); 1572 1573 /* Send ACK now, if this read freed lots of space 1574 * in our buffer. Certainly, new_window is new window. 1575 * We can advertise it now, if it is not less than current one. 1576 * "Lots" means "at least twice" here. 1577 */ 1578 if (new_window && new_window >= 2 * rcv_window_now) 1579 time_to_ack = true; 1580 } 1581 } 1582 if (time_to_ack) 1583 tcp_send_ack(sk); 1584 } 1585 1586 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1587 { 1588 struct sk_buff *skb; 1589 u32 offset; 1590 1591 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) { 1592 offset = seq - TCP_SKB_CB(skb)->seq; 1593 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 1594 pr_err_once("%s: found a SYN, please report !\n", __func__); 1595 offset--; 1596 } 1597 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) { 1598 *off = offset; 1599 return skb; 1600 } 1601 /* This looks weird, but this can happen if TCP collapsing 1602 * splitted a fat GRO packet, while we released socket lock 1603 * in skb_splice_bits() 1604 */ 1605 sk_eat_skb(sk, skb); 1606 } 1607 return NULL; 1608 } 1609 1610 /* 1611 * This routine provides an alternative to tcp_recvmsg() for routines 1612 * that would like to handle copying from skbuffs directly in 'sendfile' 1613 * fashion. 1614 * Note: 1615 * - It is assumed that the socket was locked by the caller. 1616 * - The routine does not block. 1617 * - At present, there is no support for reading OOB data 1618 * or for 'peeking' the socket using this routine 1619 * (although both would be easy to implement). 1620 */ 1621 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1622 sk_read_actor_t recv_actor) 1623 { 1624 struct sk_buff *skb; 1625 struct tcp_sock *tp = tcp_sk(sk); 1626 u32 seq = tp->copied_seq; 1627 u32 offset; 1628 int copied = 0; 1629 1630 if (sk->sk_state == TCP_LISTEN) 1631 return -ENOTCONN; 1632 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1633 if (offset < skb->len) { 1634 int used; 1635 size_t len; 1636 1637 len = skb->len - offset; 1638 /* Stop reading if we hit a patch of urgent data */ 1639 if (tp->urg_data) { 1640 u32 urg_offset = tp->urg_seq - seq; 1641 if (urg_offset < len) 1642 len = urg_offset; 1643 if (!len) 1644 break; 1645 } 1646 used = recv_actor(desc, skb, offset, len); 1647 if (used <= 0) { 1648 if (!copied) 1649 copied = used; 1650 break; 1651 } else if (used <= len) { 1652 seq += used; 1653 copied += used; 1654 offset += used; 1655 } 1656 /* If recv_actor drops the lock (e.g. TCP splice 1657 * receive) the skb pointer might be invalid when 1658 * getting here: tcp_collapse might have deleted it 1659 * while aggregating skbs from the socket queue. 1660 */ 1661 skb = tcp_recv_skb(sk, seq - 1, &offset); 1662 if (!skb) 1663 break; 1664 /* TCP coalescing might have appended data to the skb. 1665 * Try to splice more frags 1666 */ 1667 if (offset + 1 != skb->len) 1668 continue; 1669 } 1670 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { 1671 sk_eat_skb(sk, skb); 1672 ++seq; 1673 break; 1674 } 1675 sk_eat_skb(sk, skb); 1676 if (!desc->count) 1677 break; 1678 WRITE_ONCE(tp->copied_seq, seq); 1679 } 1680 WRITE_ONCE(tp->copied_seq, seq); 1681 1682 tcp_rcv_space_adjust(sk); 1683 1684 /* Clean up data we have read: This will do ACK frames. */ 1685 if (copied > 0) { 1686 tcp_recv_skb(sk, seq, &offset); 1687 tcp_cleanup_rbuf(sk, copied); 1688 } 1689 return copied; 1690 } 1691 EXPORT_SYMBOL(tcp_read_sock); 1692 1693 int tcp_peek_len(struct socket *sock) 1694 { 1695 return tcp_inq(sock->sk); 1696 } 1697 EXPORT_SYMBOL(tcp_peek_len); 1698 1699 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */ 1700 int tcp_set_rcvlowat(struct sock *sk, int val) 1701 { 1702 int cap; 1703 1704 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK) 1705 cap = sk->sk_rcvbuf >> 1; 1706 else 1707 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1; 1708 val = min(val, cap); 1709 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1); 1710 1711 /* Check if we need to signal EPOLLIN right now */ 1712 tcp_data_ready(sk); 1713 1714 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK) 1715 return 0; 1716 1717 val <<= 1; 1718 if (val > sk->sk_rcvbuf) { 1719 WRITE_ONCE(sk->sk_rcvbuf, val); 1720 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val); 1721 } 1722 return 0; 1723 } 1724 EXPORT_SYMBOL(tcp_set_rcvlowat); 1725 1726 #ifdef CONFIG_MMU 1727 static const struct vm_operations_struct tcp_vm_ops = { 1728 }; 1729 1730 int tcp_mmap(struct file *file, struct socket *sock, 1731 struct vm_area_struct *vma) 1732 { 1733 if (vma->vm_flags & (VM_WRITE | VM_EXEC)) 1734 return -EPERM; 1735 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC); 1736 1737 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */ 1738 vma->vm_flags |= VM_MIXEDMAP; 1739 1740 vma->vm_ops = &tcp_vm_ops; 1741 return 0; 1742 } 1743 EXPORT_SYMBOL(tcp_mmap); 1744 1745 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma, 1746 struct page **pages, 1747 unsigned long pages_to_map, 1748 unsigned long *insert_addr, 1749 u32 *length_with_pending, 1750 u32 *seq, 1751 struct tcp_zerocopy_receive *zc) 1752 { 1753 unsigned long pages_remaining = pages_to_map; 1754 int bytes_mapped; 1755 int ret; 1756 1757 ret = vm_insert_pages(vma, *insert_addr, pages, &pages_remaining); 1758 bytes_mapped = PAGE_SIZE * (pages_to_map - pages_remaining); 1759 /* Even if vm_insert_pages fails, it may have partially succeeded in 1760 * mapping (some but not all of the pages). 1761 */ 1762 *seq += bytes_mapped; 1763 *insert_addr += bytes_mapped; 1764 if (ret) { 1765 /* But if vm_insert_pages did fail, we have to unroll some state 1766 * we speculatively touched before. 1767 */ 1768 const int bytes_not_mapped = PAGE_SIZE * pages_remaining; 1769 *length_with_pending -= bytes_not_mapped; 1770 zc->recv_skip_hint += bytes_not_mapped; 1771 } 1772 return ret; 1773 } 1774 1775 static int tcp_zerocopy_receive(struct sock *sk, 1776 struct tcp_zerocopy_receive *zc) 1777 { 1778 unsigned long address = (unsigned long)zc->address; 1779 u32 length = 0, seq, offset, zap_len; 1780 #define PAGE_BATCH_SIZE 8 1781 struct page *pages[PAGE_BATCH_SIZE]; 1782 const skb_frag_t *frags = NULL; 1783 struct vm_area_struct *vma; 1784 struct sk_buff *skb = NULL; 1785 unsigned long pg_idx = 0; 1786 unsigned long curr_addr; 1787 struct tcp_sock *tp; 1788 int inq; 1789 int ret; 1790 1791 if (address & (PAGE_SIZE - 1) || address != zc->address) 1792 return -EINVAL; 1793 1794 if (sk->sk_state == TCP_LISTEN) 1795 return -ENOTCONN; 1796 1797 sock_rps_record_flow(sk); 1798 1799 tp = tcp_sk(sk); 1800 1801 mmap_read_lock(current->mm); 1802 1803 vma = find_vma(current->mm, address); 1804 if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops) { 1805 mmap_read_unlock(current->mm); 1806 return -EINVAL; 1807 } 1808 zc->length = min_t(unsigned long, zc->length, vma->vm_end - address); 1809 1810 seq = tp->copied_seq; 1811 inq = tcp_inq(sk); 1812 zc->length = min_t(u32, zc->length, inq); 1813 zap_len = zc->length & ~(PAGE_SIZE - 1); 1814 if (zap_len) { 1815 zap_page_range(vma, address, zap_len); 1816 zc->recv_skip_hint = 0; 1817 } else { 1818 zc->recv_skip_hint = zc->length; 1819 } 1820 ret = 0; 1821 curr_addr = address; 1822 while (length + PAGE_SIZE <= zc->length) { 1823 if (zc->recv_skip_hint < PAGE_SIZE) { 1824 /* If we're here, finish the current batch. */ 1825 if (pg_idx) { 1826 ret = tcp_zerocopy_vm_insert_batch(vma, pages, 1827 pg_idx, 1828 &curr_addr, 1829 &length, 1830 &seq, zc); 1831 if (ret) 1832 goto out; 1833 pg_idx = 0; 1834 } 1835 if (skb) { 1836 if (zc->recv_skip_hint > 0) 1837 break; 1838 skb = skb->next; 1839 offset = seq - TCP_SKB_CB(skb)->seq; 1840 } else { 1841 skb = tcp_recv_skb(sk, seq, &offset); 1842 } 1843 zc->recv_skip_hint = skb->len - offset; 1844 offset -= skb_headlen(skb); 1845 if ((int)offset < 0 || skb_has_frag_list(skb)) 1846 break; 1847 frags = skb_shinfo(skb)->frags; 1848 while (offset) { 1849 if (skb_frag_size(frags) > offset) 1850 goto out; 1851 offset -= skb_frag_size(frags); 1852 frags++; 1853 } 1854 } 1855 if (skb_frag_size(frags) != PAGE_SIZE || skb_frag_off(frags)) { 1856 int remaining = zc->recv_skip_hint; 1857 1858 while (remaining && (skb_frag_size(frags) != PAGE_SIZE || 1859 skb_frag_off(frags))) { 1860 remaining -= skb_frag_size(frags); 1861 frags++; 1862 } 1863 zc->recv_skip_hint -= remaining; 1864 break; 1865 } 1866 pages[pg_idx] = skb_frag_page(frags); 1867 pg_idx++; 1868 length += PAGE_SIZE; 1869 zc->recv_skip_hint -= PAGE_SIZE; 1870 frags++; 1871 if (pg_idx == PAGE_BATCH_SIZE) { 1872 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pg_idx, 1873 &curr_addr, &length, 1874 &seq, zc); 1875 if (ret) 1876 goto out; 1877 pg_idx = 0; 1878 } 1879 } 1880 if (pg_idx) { 1881 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pg_idx, 1882 &curr_addr, &length, &seq, 1883 zc); 1884 } 1885 out: 1886 mmap_read_unlock(current->mm); 1887 if (length) { 1888 WRITE_ONCE(tp->copied_seq, seq); 1889 tcp_rcv_space_adjust(sk); 1890 1891 /* Clean up data we have read: This will do ACK frames. */ 1892 tcp_recv_skb(sk, seq, &offset); 1893 tcp_cleanup_rbuf(sk, length); 1894 ret = 0; 1895 if (length == zc->length) 1896 zc->recv_skip_hint = 0; 1897 } else { 1898 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE)) 1899 ret = -EIO; 1900 } 1901 zc->length = length; 1902 return ret; 1903 } 1904 #endif 1905 1906 static void tcp_update_recv_tstamps(struct sk_buff *skb, 1907 struct scm_timestamping_internal *tss) 1908 { 1909 if (skb->tstamp) 1910 tss->ts[0] = ktime_to_timespec64(skb->tstamp); 1911 else 1912 tss->ts[0] = (struct timespec64) {0}; 1913 1914 if (skb_hwtstamps(skb)->hwtstamp) 1915 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp); 1916 else 1917 tss->ts[2] = (struct timespec64) {0}; 1918 } 1919 1920 /* Similar to __sock_recv_timestamp, but does not require an skb */ 1921 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk, 1922 struct scm_timestamping_internal *tss) 1923 { 1924 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW); 1925 bool has_timestamping = false; 1926 1927 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) { 1928 if (sock_flag(sk, SOCK_RCVTSTAMP)) { 1929 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) { 1930 if (new_tstamp) { 1931 struct __kernel_timespec kts = { 1932 .tv_sec = tss->ts[0].tv_sec, 1933 .tv_nsec = tss->ts[0].tv_nsec, 1934 }; 1935 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW, 1936 sizeof(kts), &kts); 1937 } else { 1938 struct __kernel_old_timespec ts_old = { 1939 .tv_sec = tss->ts[0].tv_sec, 1940 .tv_nsec = tss->ts[0].tv_nsec, 1941 }; 1942 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD, 1943 sizeof(ts_old), &ts_old); 1944 } 1945 } else { 1946 if (new_tstamp) { 1947 struct __kernel_sock_timeval stv = { 1948 .tv_sec = tss->ts[0].tv_sec, 1949 .tv_usec = tss->ts[0].tv_nsec / 1000, 1950 }; 1951 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW, 1952 sizeof(stv), &stv); 1953 } else { 1954 struct __kernel_old_timeval tv = { 1955 .tv_sec = tss->ts[0].tv_sec, 1956 .tv_usec = tss->ts[0].tv_nsec / 1000, 1957 }; 1958 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD, 1959 sizeof(tv), &tv); 1960 } 1961 } 1962 } 1963 1964 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE) 1965 has_timestamping = true; 1966 else 1967 tss->ts[0] = (struct timespec64) {0}; 1968 } 1969 1970 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) { 1971 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE) 1972 has_timestamping = true; 1973 else 1974 tss->ts[2] = (struct timespec64) {0}; 1975 } 1976 1977 if (has_timestamping) { 1978 tss->ts[1] = (struct timespec64) {0}; 1979 if (sock_flag(sk, SOCK_TSTAMP_NEW)) 1980 put_cmsg_scm_timestamping64(msg, tss); 1981 else 1982 put_cmsg_scm_timestamping(msg, tss); 1983 } 1984 } 1985 1986 static int tcp_inq_hint(struct sock *sk) 1987 { 1988 const struct tcp_sock *tp = tcp_sk(sk); 1989 u32 copied_seq = READ_ONCE(tp->copied_seq); 1990 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt); 1991 int inq; 1992 1993 inq = rcv_nxt - copied_seq; 1994 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) { 1995 lock_sock(sk); 1996 inq = tp->rcv_nxt - tp->copied_seq; 1997 release_sock(sk); 1998 } 1999 /* After receiving a FIN, tell the user-space to continue reading 2000 * by returning a non-zero inq. 2001 */ 2002 if (inq == 0 && sock_flag(sk, SOCK_DONE)) 2003 inq = 1; 2004 return inq; 2005 } 2006 2007 /* 2008 * This routine copies from a sock struct into the user buffer. 2009 * 2010 * Technical note: in 2.3 we work on _locked_ socket, so that 2011 * tricks with *seq access order and skb->users are not required. 2012 * Probably, code can be easily improved even more. 2013 */ 2014 2015 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock, 2016 int flags, int *addr_len) 2017 { 2018 struct tcp_sock *tp = tcp_sk(sk); 2019 int copied = 0; 2020 u32 peek_seq; 2021 u32 *seq; 2022 unsigned long used; 2023 int err, inq; 2024 int target; /* Read at least this many bytes */ 2025 long timeo; 2026 struct sk_buff *skb, *last; 2027 u32 urg_hole = 0; 2028 struct scm_timestamping_internal tss; 2029 int cmsg_flags; 2030 2031 if (unlikely(flags & MSG_ERRQUEUE)) 2032 return inet_recv_error(sk, msg, len, addr_len); 2033 2034 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) && 2035 (sk->sk_state == TCP_ESTABLISHED)) 2036 sk_busy_loop(sk, nonblock); 2037 2038 lock_sock(sk); 2039 2040 err = -ENOTCONN; 2041 if (sk->sk_state == TCP_LISTEN) 2042 goto out; 2043 2044 cmsg_flags = tp->recvmsg_inq ? 1 : 0; 2045 timeo = sock_rcvtimeo(sk, nonblock); 2046 2047 /* Urgent data needs to be handled specially. */ 2048 if (flags & MSG_OOB) 2049 goto recv_urg; 2050 2051 if (unlikely(tp->repair)) { 2052 err = -EPERM; 2053 if (!(flags & MSG_PEEK)) 2054 goto out; 2055 2056 if (tp->repair_queue == TCP_SEND_QUEUE) 2057 goto recv_sndq; 2058 2059 err = -EINVAL; 2060 if (tp->repair_queue == TCP_NO_QUEUE) 2061 goto out; 2062 2063 /* 'common' recv queue MSG_PEEK-ing */ 2064 } 2065 2066 seq = &tp->copied_seq; 2067 if (flags & MSG_PEEK) { 2068 peek_seq = tp->copied_seq; 2069 seq = &peek_seq; 2070 } 2071 2072 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2073 2074 do { 2075 u32 offset; 2076 2077 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 2078 if (tp->urg_data && tp->urg_seq == *seq) { 2079 if (copied) 2080 break; 2081 if (signal_pending(current)) { 2082 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 2083 break; 2084 } 2085 } 2086 2087 /* Next get a buffer. */ 2088 2089 last = skb_peek_tail(&sk->sk_receive_queue); 2090 skb_queue_walk(&sk->sk_receive_queue, skb) { 2091 last = skb; 2092 /* Now that we have two receive queues this 2093 * shouldn't happen. 2094 */ 2095 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 2096 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n", 2097 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 2098 flags)) 2099 break; 2100 2101 offset = *seq - TCP_SKB_CB(skb)->seq; 2102 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 2103 pr_err_once("%s: found a SYN, please report !\n", __func__); 2104 offset--; 2105 } 2106 if (offset < skb->len) 2107 goto found_ok_skb; 2108 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2109 goto found_fin_ok; 2110 WARN(!(flags & MSG_PEEK), 2111 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n", 2112 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 2113 } 2114 2115 /* Well, if we have backlog, try to process it now yet. */ 2116 2117 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail)) 2118 break; 2119 2120 if (copied) { 2121 if (sk->sk_err || 2122 sk->sk_state == TCP_CLOSE || 2123 (sk->sk_shutdown & RCV_SHUTDOWN) || 2124 !timeo || 2125 signal_pending(current)) 2126 break; 2127 } else { 2128 if (sock_flag(sk, SOCK_DONE)) 2129 break; 2130 2131 if (sk->sk_err) { 2132 copied = sock_error(sk); 2133 break; 2134 } 2135 2136 if (sk->sk_shutdown & RCV_SHUTDOWN) 2137 break; 2138 2139 if (sk->sk_state == TCP_CLOSE) { 2140 /* This occurs when user tries to read 2141 * from never connected socket. 2142 */ 2143 copied = -ENOTCONN; 2144 break; 2145 } 2146 2147 if (!timeo) { 2148 copied = -EAGAIN; 2149 break; 2150 } 2151 2152 if (signal_pending(current)) { 2153 copied = sock_intr_errno(timeo); 2154 break; 2155 } 2156 } 2157 2158 tcp_cleanup_rbuf(sk, copied); 2159 2160 if (copied >= target) { 2161 /* Do not sleep, just process backlog. */ 2162 release_sock(sk); 2163 lock_sock(sk); 2164 } else { 2165 sk_wait_data(sk, &timeo, last); 2166 } 2167 2168 if ((flags & MSG_PEEK) && 2169 (peek_seq - copied - urg_hole != tp->copied_seq)) { 2170 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n", 2171 current->comm, 2172 task_pid_nr(current)); 2173 peek_seq = tp->copied_seq; 2174 } 2175 continue; 2176 2177 found_ok_skb: 2178 /* Ok so how much can we use? */ 2179 used = skb->len - offset; 2180 if (len < used) 2181 used = len; 2182 2183 /* Do we have urgent data here? */ 2184 if (tp->urg_data) { 2185 u32 urg_offset = tp->urg_seq - *seq; 2186 if (urg_offset < used) { 2187 if (!urg_offset) { 2188 if (!sock_flag(sk, SOCK_URGINLINE)) { 2189 WRITE_ONCE(*seq, *seq + 1); 2190 urg_hole++; 2191 offset++; 2192 used--; 2193 if (!used) 2194 goto skip_copy; 2195 } 2196 } else 2197 used = urg_offset; 2198 } 2199 } 2200 2201 if (!(flags & MSG_TRUNC)) { 2202 err = skb_copy_datagram_msg(skb, offset, msg, used); 2203 if (err) { 2204 /* Exception. Bailout! */ 2205 if (!copied) 2206 copied = -EFAULT; 2207 break; 2208 } 2209 } 2210 2211 WRITE_ONCE(*seq, *seq + used); 2212 copied += used; 2213 len -= used; 2214 2215 tcp_rcv_space_adjust(sk); 2216 2217 skip_copy: 2218 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 2219 tp->urg_data = 0; 2220 tcp_fast_path_check(sk); 2221 } 2222 2223 if (TCP_SKB_CB(skb)->has_rxtstamp) { 2224 tcp_update_recv_tstamps(skb, &tss); 2225 cmsg_flags |= 2; 2226 } 2227 2228 if (used + offset < skb->len) 2229 continue; 2230 2231 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2232 goto found_fin_ok; 2233 if (!(flags & MSG_PEEK)) 2234 sk_eat_skb(sk, skb); 2235 continue; 2236 2237 found_fin_ok: 2238 /* Process the FIN. */ 2239 WRITE_ONCE(*seq, *seq + 1); 2240 if (!(flags & MSG_PEEK)) 2241 sk_eat_skb(sk, skb); 2242 break; 2243 } while (len > 0); 2244 2245 /* According to UNIX98, msg_name/msg_namelen are ignored 2246 * on connected socket. I was just happy when found this 8) --ANK 2247 */ 2248 2249 /* Clean up data we have read: This will do ACK frames. */ 2250 tcp_cleanup_rbuf(sk, copied); 2251 2252 release_sock(sk); 2253 2254 if (cmsg_flags) { 2255 if (cmsg_flags & 2) 2256 tcp_recv_timestamp(msg, sk, &tss); 2257 if (cmsg_flags & 1) { 2258 inq = tcp_inq_hint(sk); 2259 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq); 2260 } 2261 } 2262 2263 return copied; 2264 2265 out: 2266 release_sock(sk); 2267 return err; 2268 2269 recv_urg: 2270 err = tcp_recv_urg(sk, msg, len, flags); 2271 goto out; 2272 2273 recv_sndq: 2274 err = tcp_peek_sndq(sk, msg, len); 2275 goto out; 2276 } 2277 EXPORT_SYMBOL(tcp_recvmsg); 2278 2279 void tcp_set_state(struct sock *sk, int state) 2280 { 2281 int oldstate = sk->sk_state; 2282 2283 /* We defined a new enum for TCP states that are exported in BPF 2284 * so as not force the internal TCP states to be frozen. The 2285 * following checks will detect if an internal state value ever 2286 * differs from the BPF value. If this ever happens, then we will 2287 * need to remap the internal value to the BPF value before calling 2288 * tcp_call_bpf_2arg. 2289 */ 2290 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED); 2291 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT); 2292 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV); 2293 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1); 2294 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2); 2295 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT); 2296 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE); 2297 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT); 2298 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK); 2299 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN); 2300 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING); 2301 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV); 2302 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES); 2303 2304 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG)) 2305 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state); 2306 2307 switch (state) { 2308 case TCP_ESTABLISHED: 2309 if (oldstate != TCP_ESTABLISHED) 2310 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 2311 break; 2312 2313 case TCP_CLOSE: 2314 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 2315 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 2316 2317 sk->sk_prot->unhash(sk); 2318 if (inet_csk(sk)->icsk_bind_hash && 2319 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 2320 inet_put_port(sk); 2321 fallthrough; 2322 default: 2323 if (oldstate == TCP_ESTABLISHED) 2324 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 2325 } 2326 2327 /* Change state AFTER socket is unhashed to avoid closed 2328 * socket sitting in hash tables. 2329 */ 2330 inet_sk_state_store(sk, state); 2331 } 2332 EXPORT_SYMBOL_GPL(tcp_set_state); 2333 2334 /* 2335 * State processing on a close. This implements the state shift for 2336 * sending our FIN frame. Note that we only send a FIN for some 2337 * states. A shutdown() may have already sent the FIN, or we may be 2338 * closed. 2339 */ 2340 2341 static const unsigned char new_state[16] = { 2342 /* current state: new state: action: */ 2343 [0 /* (Invalid) */] = TCP_CLOSE, 2344 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2345 [TCP_SYN_SENT] = TCP_CLOSE, 2346 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2347 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 2348 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 2349 [TCP_TIME_WAIT] = TCP_CLOSE, 2350 [TCP_CLOSE] = TCP_CLOSE, 2351 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 2352 [TCP_LAST_ACK] = TCP_LAST_ACK, 2353 [TCP_LISTEN] = TCP_CLOSE, 2354 [TCP_CLOSING] = TCP_CLOSING, 2355 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 2356 }; 2357 2358 static int tcp_close_state(struct sock *sk) 2359 { 2360 int next = (int)new_state[sk->sk_state]; 2361 int ns = next & TCP_STATE_MASK; 2362 2363 tcp_set_state(sk, ns); 2364 2365 return next & TCP_ACTION_FIN; 2366 } 2367 2368 /* 2369 * Shutdown the sending side of a connection. Much like close except 2370 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 2371 */ 2372 2373 void tcp_shutdown(struct sock *sk, int how) 2374 { 2375 /* We need to grab some memory, and put together a FIN, 2376 * and then put it into the queue to be sent. 2377 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 2378 */ 2379 if (!(how & SEND_SHUTDOWN)) 2380 return; 2381 2382 /* If we've already sent a FIN, or it's a closed state, skip this. */ 2383 if ((1 << sk->sk_state) & 2384 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 2385 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 2386 /* Clear out any half completed packets. FIN if needed. */ 2387 if (tcp_close_state(sk)) 2388 tcp_send_fin(sk); 2389 } 2390 } 2391 EXPORT_SYMBOL(tcp_shutdown); 2392 2393 bool tcp_check_oom(struct sock *sk, int shift) 2394 { 2395 bool too_many_orphans, out_of_socket_memory; 2396 2397 too_many_orphans = tcp_too_many_orphans(sk, shift); 2398 out_of_socket_memory = tcp_out_of_memory(sk); 2399 2400 if (too_many_orphans) 2401 net_info_ratelimited("too many orphaned sockets\n"); 2402 if (out_of_socket_memory) 2403 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n"); 2404 return too_many_orphans || out_of_socket_memory; 2405 } 2406 2407 void tcp_close(struct sock *sk, long timeout) 2408 { 2409 struct sk_buff *skb; 2410 int data_was_unread = 0; 2411 int state; 2412 2413 lock_sock(sk); 2414 sk->sk_shutdown = SHUTDOWN_MASK; 2415 2416 if (sk->sk_state == TCP_LISTEN) { 2417 tcp_set_state(sk, TCP_CLOSE); 2418 2419 /* Special case. */ 2420 inet_csk_listen_stop(sk); 2421 2422 goto adjudge_to_death; 2423 } 2424 2425 /* We need to flush the recv. buffs. We do this only on the 2426 * descriptor close, not protocol-sourced closes, because the 2427 * reader process may not have drained the data yet! 2428 */ 2429 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 2430 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq; 2431 2432 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2433 len--; 2434 data_was_unread += len; 2435 __kfree_skb(skb); 2436 } 2437 2438 sk_mem_reclaim(sk); 2439 2440 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 2441 if (sk->sk_state == TCP_CLOSE) 2442 goto adjudge_to_death; 2443 2444 /* As outlined in RFC 2525, section 2.17, we send a RST here because 2445 * data was lost. To witness the awful effects of the old behavior of 2446 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 2447 * GET in an FTP client, suspend the process, wait for the client to 2448 * advertise a zero window, then kill -9 the FTP client, wheee... 2449 * Note: timeout is always zero in such a case. 2450 */ 2451 if (unlikely(tcp_sk(sk)->repair)) { 2452 sk->sk_prot->disconnect(sk, 0); 2453 } else if (data_was_unread) { 2454 /* Unread data was tossed, zap the connection. */ 2455 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 2456 tcp_set_state(sk, TCP_CLOSE); 2457 tcp_send_active_reset(sk, sk->sk_allocation); 2458 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 2459 /* Check zero linger _after_ checking for unread data. */ 2460 sk->sk_prot->disconnect(sk, 0); 2461 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 2462 } else if (tcp_close_state(sk)) { 2463 /* We FIN if the application ate all the data before 2464 * zapping the connection. 2465 */ 2466 2467 /* RED-PEN. Formally speaking, we have broken TCP state 2468 * machine. State transitions: 2469 * 2470 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 2471 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 2472 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 2473 * 2474 * are legal only when FIN has been sent (i.e. in window), 2475 * rather than queued out of window. Purists blame. 2476 * 2477 * F.e. "RFC state" is ESTABLISHED, 2478 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 2479 * 2480 * The visible declinations are that sometimes 2481 * we enter time-wait state, when it is not required really 2482 * (harmless), do not send active resets, when they are 2483 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 2484 * they look as CLOSING or LAST_ACK for Linux) 2485 * Probably, I missed some more holelets. 2486 * --ANK 2487 * XXX (TFO) - To start off we don't support SYN+ACK+FIN 2488 * in a single packet! (May consider it later but will 2489 * probably need API support or TCP_CORK SYN-ACK until 2490 * data is written and socket is closed.) 2491 */ 2492 tcp_send_fin(sk); 2493 } 2494 2495 sk_stream_wait_close(sk, timeout); 2496 2497 adjudge_to_death: 2498 state = sk->sk_state; 2499 sock_hold(sk); 2500 sock_orphan(sk); 2501 2502 local_bh_disable(); 2503 bh_lock_sock(sk); 2504 /* remove backlog if any, without releasing ownership. */ 2505 __release_sock(sk); 2506 2507 percpu_counter_inc(sk->sk_prot->orphan_count); 2508 2509 /* Have we already been destroyed by a softirq or backlog? */ 2510 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 2511 goto out; 2512 2513 /* This is a (useful) BSD violating of the RFC. There is a 2514 * problem with TCP as specified in that the other end could 2515 * keep a socket open forever with no application left this end. 2516 * We use a 1 minute timeout (about the same as BSD) then kill 2517 * our end. If they send after that then tough - BUT: long enough 2518 * that we won't make the old 4*rto = almost no time - whoops 2519 * reset mistake. 2520 * 2521 * Nope, it was not mistake. It is really desired behaviour 2522 * f.e. on http servers, when such sockets are useless, but 2523 * consume significant resources. Let's do it with special 2524 * linger2 option. --ANK 2525 */ 2526 2527 if (sk->sk_state == TCP_FIN_WAIT2) { 2528 struct tcp_sock *tp = tcp_sk(sk); 2529 if (tp->linger2 < 0) { 2530 tcp_set_state(sk, TCP_CLOSE); 2531 tcp_send_active_reset(sk, GFP_ATOMIC); 2532 __NET_INC_STATS(sock_net(sk), 2533 LINUX_MIB_TCPABORTONLINGER); 2534 } else { 2535 const int tmo = tcp_fin_time(sk); 2536 2537 if (tmo > TCP_TIMEWAIT_LEN) { 2538 inet_csk_reset_keepalive_timer(sk, 2539 tmo - TCP_TIMEWAIT_LEN); 2540 } else { 2541 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2542 goto out; 2543 } 2544 } 2545 } 2546 if (sk->sk_state != TCP_CLOSE) { 2547 sk_mem_reclaim(sk); 2548 if (tcp_check_oom(sk, 0)) { 2549 tcp_set_state(sk, TCP_CLOSE); 2550 tcp_send_active_reset(sk, GFP_ATOMIC); 2551 __NET_INC_STATS(sock_net(sk), 2552 LINUX_MIB_TCPABORTONMEMORY); 2553 } else if (!check_net(sock_net(sk))) { 2554 /* Not possible to send reset; just close */ 2555 tcp_set_state(sk, TCP_CLOSE); 2556 } 2557 } 2558 2559 if (sk->sk_state == TCP_CLOSE) { 2560 struct request_sock *req; 2561 2562 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 2563 lockdep_sock_is_held(sk)); 2564 /* We could get here with a non-NULL req if the socket is 2565 * aborted (e.g., closed with unread data) before 3WHS 2566 * finishes. 2567 */ 2568 if (req) 2569 reqsk_fastopen_remove(sk, req, false); 2570 inet_csk_destroy_sock(sk); 2571 } 2572 /* Otherwise, socket is reprieved until protocol close. */ 2573 2574 out: 2575 bh_unlock_sock(sk); 2576 local_bh_enable(); 2577 release_sock(sk); 2578 sock_put(sk); 2579 } 2580 EXPORT_SYMBOL(tcp_close); 2581 2582 /* These states need RST on ABORT according to RFC793 */ 2583 2584 static inline bool tcp_need_reset(int state) 2585 { 2586 return (1 << state) & 2587 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2588 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2589 } 2590 2591 static void tcp_rtx_queue_purge(struct sock *sk) 2592 { 2593 struct rb_node *p = rb_first(&sk->tcp_rtx_queue); 2594 2595 tcp_sk(sk)->highest_sack = NULL; 2596 while (p) { 2597 struct sk_buff *skb = rb_to_skb(p); 2598 2599 p = rb_next(p); 2600 /* Since we are deleting whole queue, no need to 2601 * list_del(&skb->tcp_tsorted_anchor) 2602 */ 2603 tcp_rtx_queue_unlink(skb, sk); 2604 sk_wmem_free_skb(sk, skb); 2605 } 2606 } 2607 2608 void tcp_write_queue_purge(struct sock *sk) 2609 { 2610 struct sk_buff *skb; 2611 2612 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 2613 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) { 2614 tcp_skb_tsorted_anchor_cleanup(skb); 2615 sk_wmem_free_skb(sk, skb); 2616 } 2617 tcp_rtx_queue_purge(sk); 2618 skb = sk->sk_tx_skb_cache; 2619 if (skb) { 2620 __kfree_skb(skb); 2621 sk->sk_tx_skb_cache = NULL; 2622 } 2623 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue); 2624 sk_mem_reclaim(sk); 2625 tcp_clear_all_retrans_hints(tcp_sk(sk)); 2626 tcp_sk(sk)->packets_out = 0; 2627 inet_csk(sk)->icsk_backoff = 0; 2628 } 2629 2630 int tcp_disconnect(struct sock *sk, int flags) 2631 { 2632 struct inet_sock *inet = inet_sk(sk); 2633 struct inet_connection_sock *icsk = inet_csk(sk); 2634 struct tcp_sock *tp = tcp_sk(sk); 2635 int old_state = sk->sk_state; 2636 u32 seq; 2637 2638 if (old_state != TCP_CLOSE) 2639 tcp_set_state(sk, TCP_CLOSE); 2640 2641 /* ABORT function of RFC793 */ 2642 if (old_state == TCP_LISTEN) { 2643 inet_csk_listen_stop(sk); 2644 } else if (unlikely(tp->repair)) { 2645 sk->sk_err = ECONNABORTED; 2646 } else if (tcp_need_reset(old_state) || 2647 (tp->snd_nxt != tp->write_seq && 2648 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2649 /* The last check adjusts for discrepancy of Linux wrt. RFC 2650 * states 2651 */ 2652 tcp_send_active_reset(sk, gfp_any()); 2653 sk->sk_err = ECONNRESET; 2654 } else if (old_state == TCP_SYN_SENT) 2655 sk->sk_err = ECONNRESET; 2656 2657 tcp_clear_xmit_timers(sk); 2658 __skb_queue_purge(&sk->sk_receive_queue); 2659 if (sk->sk_rx_skb_cache) { 2660 __kfree_skb(sk->sk_rx_skb_cache); 2661 sk->sk_rx_skb_cache = NULL; 2662 } 2663 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 2664 tp->urg_data = 0; 2665 tcp_write_queue_purge(sk); 2666 tcp_fastopen_active_disable_ofo_check(sk); 2667 skb_rbtree_purge(&tp->out_of_order_queue); 2668 2669 inet->inet_dport = 0; 2670 2671 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2672 inet_reset_saddr(sk); 2673 2674 sk->sk_shutdown = 0; 2675 sock_reset_flag(sk, SOCK_DONE); 2676 tp->srtt_us = 0; 2677 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 2678 tp->rcv_rtt_last_tsecr = 0; 2679 2680 seq = tp->write_seq + tp->max_window + 2; 2681 if (!seq) 2682 seq = 1; 2683 WRITE_ONCE(tp->write_seq, seq); 2684 2685 icsk->icsk_backoff = 0; 2686 icsk->icsk_probes_out = 0; 2687 icsk->icsk_rto = TCP_TIMEOUT_INIT; 2688 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2689 tp->snd_cwnd = TCP_INIT_CWND; 2690 tp->snd_cwnd_cnt = 0; 2691 tp->window_clamp = 0; 2692 tp->delivered = 0; 2693 tp->delivered_ce = 0; 2694 if (icsk->icsk_ca_ops->release) 2695 icsk->icsk_ca_ops->release(sk); 2696 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv)); 2697 tcp_set_ca_state(sk, TCP_CA_Open); 2698 tp->is_sack_reneg = 0; 2699 tcp_clear_retrans(tp); 2700 tp->total_retrans = 0; 2701 inet_csk_delack_init(sk); 2702 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0 2703 * issue in __tcp_select_window() 2704 */ 2705 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS; 2706 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2707 __sk_dst_reset(sk); 2708 dst_release(sk->sk_rx_dst); 2709 sk->sk_rx_dst = NULL; 2710 tcp_saved_syn_free(tp); 2711 tp->compressed_ack = 0; 2712 tp->segs_in = 0; 2713 tp->segs_out = 0; 2714 tp->bytes_sent = 0; 2715 tp->bytes_acked = 0; 2716 tp->bytes_received = 0; 2717 tp->bytes_retrans = 0; 2718 tp->data_segs_in = 0; 2719 tp->data_segs_out = 0; 2720 tp->duplicate_sack[0].start_seq = 0; 2721 tp->duplicate_sack[0].end_seq = 0; 2722 tp->dsack_dups = 0; 2723 tp->reord_seen = 0; 2724 tp->retrans_out = 0; 2725 tp->sacked_out = 0; 2726 tp->tlp_high_seq = 0; 2727 tp->last_oow_ack_time = 0; 2728 /* There's a bubble in the pipe until at least the first ACK. */ 2729 tp->app_limited = ~0U; 2730 tp->rack.mstamp = 0; 2731 tp->rack.advanced = 0; 2732 tp->rack.reo_wnd_steps = 1; 2733 tp->rack.last_delivered = 0; 2734 tp->rack.reo_wnd_persist = 0; 2735 tp->rack.dsack_seen = 0; 2736 tp->syn_data_acked = 0; 2737 tp->rx_opt.saw_tstamp = 0; 2738 tp->rx_opt.dsack = 0; 2739 tp->rx_opt.num_sacks = 0; 2740 tp->rcv_ooopack = 0; 2741 2742 2743 /* Clean up fastopen related fields */ 2744 tcp_free_fastopen_req(tp); 2745 inet->defer_connect = 0; 2746 tp->fastopen_client_fail = 0; 2747 2748 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2749 2750 if (sk->sk_frag.page) { 2751 put_page(sk->sk_frag.page); 2752 sk->sk_frag.page = NULL; 2753 sk->sk_frag.offset = 0; 2754 } 2755 2756 sk->sk_error_report(sk); 2757 return 0; 2758 } 2759 EXPORT_SYMBOL(tcp_disconnect); 2760 2761 static inline bool tcp_can_repair_sock(const struct sock *sk) 2762 { 2763 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) && 2764 (sk->sk_state != TCP_LISTEN); 2765 } 2766 2767 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len) 2768 { 2769 struct tcp_repair_window opt; 2770 2771 if (!tp->repair) 2772 return -EPERM; 2773 2774 if (len != sizeof(opt)) 2775 return -EINVAL; 2776 2777 if (copy_from_user(&opt, optbuf, sizeof(opt))) 2778 return -EFAULT; 2779 2780 if (opt.max_window < opt.snd_wnd) 2781 return -EINVAL; 2782 2783 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd)) 2784 return -EINVAL; 2785 2786 if (after(opt.rcv_wup, tp->rcv_nxt)) 2787 return -EINVAL; 2788 2789 tp->snd_wl1 = opt.snd_wl1; 2790 tp->snd_wnd = opt.snd_wnd; 2791 tp->max_window = opt.max_window; 2792 2793 tp->rcv_wnd = opt.rcv_wnd; 2794 tp->rcv_wup = opt.rcv_wup; 2795 2796 return 0; 2797 } 2798 2799 static int tcp_repair_options_est(struct sock *sk, 2800 struct tcp_repair_opt __user *optbuf, unsigned int len) 2801 { 2802 struct tcp_sock *tp = tcp_sk(sk); 2803 struct tcp_repair_opt opt; 2804 2805 while (len >= sizeof(opt)) { 2806 if (copy_from_user(&opt, optbuf, sizeof(opt))) 2807 return -EFAULT; 2808 2809 optbuf++; 2810 len -= sizeof(opt); 2811 2812 switch (opt.opt_code) { 2813 case TCPOPT_MSS: 2814 tp->rx_opt.mss_clamp = opt.opt_val; 2815 tcp_mtup_init(sk); 2816 break; 2817 case TCPOPT_WINDOW: 2818 { 2819 u16 snd_wscale = opt.opt_val & 0xFFFF; 2820 u16 rcv_wscale = opt.opt_val >> 16; 2821 2822 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE) 2823 return -EFBIG; 2824 2825 tp->rx_opt.snd_wscale = snd_wscale; 2826 tp->rx_opt.rcv_wscale = rcv_wscale; 2827 tp->rx_opt.wscale_ok = 1; 2828 } 2829 break; 2830 case TCPOPT_SACK_PERM: 2831 if (opt.opt_val != 0) 2832 return -EINVAL; 2833 2834 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 2835 break; 2836 case TCPOPT_TIMESTAMP: 2837 if (opt.opt_val != 0) 2838 return -EINVAL; 2839 2840 tp->rx_opt.tstamp_ok = 1; 2841 break; 2842 } 2843 } 2844 2845 return 0; 2846 } 2847 2848 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled); 2849 EXPORT_SYMBOL(tcp_tx_delay_enabled); 2850 2851 static void tcp_enable_tx_delay(void) 2852 { 2853 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) { 2854 static int __tcp_tx_delay_enabled = 0; 2855 2856 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) { 2857 static_branch_enable(&tcp_tx_delay_enabled); 2858 pr_info("TCP_TX_DELAY enabled\n"); 2859 } 2860 } 2861 } 2862 2863 /* When set indicates to always queue non-full frames. Later the user clears 2864 * this option and we transmit any pending partial frames in the queue. This is 2865 * meant to be used alongside sendfile() to get properly filled frames when the 2866 * user (for example) must write out headers with a write() call first and then 2867 * use sendfile to send out the data parts. 2868 * 2869 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than 2870 * TCP_NODELAY. 2871 */ 2872 static void __tcp_sock_set_cork(struct sock *sk, bool on) 2873 { 2874 struct tcp_sock *tp = tcp_sk(sk); 2875 2876 if (on) { 2877 tp->nonagle |= TCP_NAGLE_CORK; 2878 } else { 2879 tp->nonagle &= ~TCP_NAGLE_CORK; 2880 if (tp->nonagle & TCP_NAGLE_OFF) 2881 tp->nonagle |= TCP_NAGLE_PUSH; 2882 tcp_push_pending_frames(sk); 2883 } 2884 } 2885 2886 void tcp_sock_set_cork(struct sock *sk, bool on) 2887 { 2888 lock_sock(sk); 2889 __tcp_sock_set_cork(sk, on); 2890 release_sock(sk); 2891 } 2892 EXPORT_SYMBOL(tcp_sock_set_cork); 2893 2894 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is 2895 * remembered, but it is not activated until cork is cleared. 2896 * 2897 * However, when TCP_NODELAY is set we make an explicit push, which overrides 2898 * even TCP_CORK for currently queued segments. 2899 */ 2900 static void __tcp_sock_set_nodelay(struct sock *sk, bool on) 2901 { 2902 if (on) { 2903 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2904 tcp_push_pending_frames(sk); 2905 } else { 2906 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF; 2907 } 2908 } 2909 2910 void tcp_sock_set_nodelay(struct sock *sk) 2911 { 2912 lock_sock(sk); 2913 __tcp_sock_set_nodelay(sk, true); 2914 release_sock(sk); 2915 } 2916 EXPORT_SYMBOL(tcp_sock_set_nodelay); 2917 2918 static void __tcp_sock_set_quickack(struct sock *sk, int val) 2919 { 2920 if (!val) { 2921 inet_csk_enter_pingpong_mode(sk); 2922 return; 2923 } 2924 2925 inet_csk_exit_pingpong_mode(sk); 2926 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 2927 inet_csk_ack_scheduled(sk)) { 2928 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED; 2929 tcp_cleanup_rbuf(sk, 1); 2930 if (!(val & 1)) 2931 inet_csk_enter_pingpong_mode(sk); 2932 } 2933 } 2934 2935 void tcp_sock_set_quickack(struct sock *sk, int val) 2936 { 2937 lock_sock(sk); 2938 __tcp_sock_set_quickack(sk, val); 2939 release_sock(sk); 2940 } 2941 EXPORT_SYMBOL(tcp_sock_set_quickack); 2942 2943 int tcp_sock_set_syncnt(struct sock *sk, int val) 2944 { 2945 if (val < 1 || val > MAX_TCP_SYNCNT) 2946 return -EINVAL; 2947 2948 lock_sock(sk); 2949 inet_csk(sk)->icsk_syn_retries = val; 2950 release_sock(sk); 2951 return 0; 2952 } 2953 EXPORT_SYMBOL(tcp_sock_set_syncnt); 2954 2955 void tcp_sock_set_user_timeout(struct sock *sk, u32 val) 2956 { 2957 lock_sock(sk); 2958 inet_csk(sk)->icsk_user_timeout = val; 2959 release_sock(sk); 2960 } 2961 EXPORT_SYMBOL(tcp_sock_set_user_timeout); 2962 2963 static int __tcp_sock_set_keepidle(struct sock *sk, int val) 2964 { 2965 struct tcp_sock *tp = tcp_sk(sk); 2966 2967 if (val < 1 || val > MAX_TCP_KEEPIDLE) 2968 return -EINVAL; 2969 2970 tp->keepalive_time = val * HZ; 2971 if (sock_flag(sk, SOCK_KEEPOPEN) && 2972 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) { 2973 u32 elapsed = keepalive_time_elapsed(tp); 2974 2975 if (tp->keepalive_time > elapsed) 2976 elapsed = tp->keepalive_time - elapsed; 2977 else 2978 elapsed = 0; 2979 inet_csk_reset_keepalive_timer(sk, elapsed); 2980 } 2981 2982 return 0; 2983 } 2984 2985 int tcp_sock_set_keepidle(struct sock *sk, int val) 2986 { 2987 int err; 2988 2989 lock_sock(sk); 2990 err = __tcp_sock_set_keepidle(sk, val); 2991 release_sock(sk); 2992 return err; 2993 } 2994 EXPORT_SYMBOL(tcp_sock_set_keepidle); 2995 2996 int tcp_sock_set_keepintvl(struct sock *sk, int val) 2997 { 2998 if (val < 1 || val > MAX_TCP_KEEPINTVL) 2999 return -EINVAL; 3000 3001 lock_sock(sk); 3002 tcp_sk(sk)->keepalive_intvl = val * HZ; 3003 release_sock(sk); 3004 return 0; 3005 } 3006 EXPORT_SYMBOL(tcp_sock_set_keepintvl); 3007 3008 int tcp_sock_set_keepcnt(struct sock *sk, int val) 3009 { 3010 if (val < 1 || val > MAX_TCP_KEEPCNT) 3011 return -EINVAL; 3012 3013 lock_sock(sk); 3014 tcp_sk(sk)->keepalive_probes = val; 3015 release_sock(sk); 3016 return 0; 3017 } 3018 EXPORT_SYMBOL(tcp_sock_set_keepcnt); 3019 3020 /* 3021 * Socket option code for TCP. 3022 */ 3023 static int do_tcp_setsockopt(struct sock *sk, int level, 3024 int optname, char __user *optval, unsigned int optlen) 3025 { 3026 struct tcp_sock *tp = tcp_sk(sk); 3027 struct inet_connection_sock *icsk = inet_csk(sk); 3028 struct net *net = sock_net(sk); 3029 int val; 3030 int err = 0; 3031 3032 /* These are data/string values, all the others are ints */ 3033 switch (optname) { 3034 case TCP_CONGESTION: { 3035 char name[TCP_CA_NAME_MAX]; 3036 3037 if (optlen < 1) 3038 return -EINVAL; 3039 3040 val = strncpy_from_user(name, optval, 3041 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 3042 if (val < 0) 3043 return -EFAULT; 3044 name[val] = 0; 3045 3046 lock_sock(sk); 3047 err = tcp_set_congestion_control(sk, name, true, true, 3048 ns_capable(sock_net(sk)->user_ns, 3049 CAP_NET_ADMIN)); 3050 release_sock(sk); 3051 return err; 3052 } 3053 case TCP_ULP: { 3054 char name[TCP_ULP_NAME_MAX]; 3055 3056 if (optlen < 1) 3057 return -EINVAL; 3058 3059 val = strncpy_from_user(name, optval, 3060 min_t(long, TCP_ULP_NAME_MAX - 1, 3061 optlen)); 3062 if (val < 0) 3063 return -EFAULT; 3064 name[val] = 0; 3065 3066 lock_sock(sk); 3067 err = tcp_set_ulp(sk, name); 3068 release_sock(sk); 3069 return err; 3070 } 3071 case TCP_FASTOPEN_KEY: { 3072 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH]; 3073 __u8 *backup_key = NULL; 3074 3075 /* Allow a backup key as well to facilitate key rotation 3076 * First key is the active one. 3077 */ 3078 if (optlen != TCP_FASTOPEN_KEY_LENGTH && 3079 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH) 3080 return -EINVAL; 3081 3082 if (copy_from_user(key, optval, optlen)) 3083 return -EFAULT; 3084 3085 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH) 3086 backup_key = key + TCP_FASTOPEN_KEY_LENGTH; 3087 3088 return tcp_fastopen_reset_cipher(net, sk, key, backup_key); 3089 } 3090 default: 3091 /* fallthru */ 3092 break; 3093 } 3094 3095 if (optlen < sizeof(int)) 3096 return -EINVAL; 3097 3098 if (get_user(val, (int __user *)optval)) 3099 return -EFAULT; 3100 3101 lock_sock(sk); 3102 3103 switch (optname) { 3104 case TCP_MAXSEG: 3105 /* Values greater than interface MTU won't take effect. However 3106 * at the point when this call is done we typically don't yet 3107 * know which interface is going to be used 3108 */ 3109 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) { 3110 err = -EINVAL; 3111 break; 3112 } 3113 tp->rx_opt.user_mss = val; 3114 break; 3115 3116 case TCP_NODELAY: 3117 __tcp_sock_set_nodelay(sk, val); 3118 break; 3119 3120 case TCP_THIN_LINEAR_TIMEOUTS: 3121 if (val < 0 || val > 1) 3122 err = -EINVAL; 3123 else 3124 tp->thin_lto = val; 3125 break; 3126 3127 case TCP_THIN_DUPACK: 3128 if (val < 0 || val > 1) 3129 err = -EINVAL; 3130 break; 3131 3132 case TCP_REPAIR: 3133 if (!tcp_can_repair_sock(sk)) 3134 err = -EPERM; 3135 else if (val == TCP_REPAIR_ON) { 3136 tp->repair = 1; 3137 sk->sk_reuse = SK_FORCE_REUSE; 3138 tp->repair_queue = TCP_NO_QUEUE; 3139 } else if (val == TCP_REPAIR_OFF) { 3140 tp->repair = 0; 3141 sk->sk_reuse = SK_NO_REUSE; 3142 tcp_send_window_probe(sk); 3143 } else if (val == TCP_REPAIR_OFF_NO_WP) { 3144 tp->repair = 0; 3145 sk->sk_reuse = SK_NO_REUSE; 3146 } else 3147 err = -EINVAL; 3148 3149 break; 3150 3151 case TCP_REPAIR_QUEUE: 3152 if (!tp->repair) 3153 err = -EPERM; 3154 else if ((unsigned int)val < TCP_QUEUES_NR) 3155 tp->repair_queue = val; 3156 else 3157 err = -EINVAL; 3158 break; 3159 3160 case TCP_QUEUE_SEQ: 3161 if (sk->sk_state != TCP_CLOSE) 3162 err = -EPERM; 3163 else if (tp->repair_queue == TCP_SEND_QUEUE) 3164 WRITE_ONCE(tp->write_seq, val); 3165 else if (tp->repair_queue == TCP_RECV_QUEUE) { 3166 WRITE_ONCE(tp->rcv_nxt, val); 3167 WRITE_ONCE(tp->copied_seq, val); 3168 } 3169 else 3170 err = -EINVAL; 3171 break; 3172 3173 case TCP_REPAIR_OPTIONS: 3174 if (!tp->repair) 3175 err = -EINVAL; 3176 else if (sk->sk_state == TCP_ESTABLISHED) 3177 err = tcp_repair_options_est(sk, 3178 (struct tcp_repair_opt __user *)optval, 3179 optlen); 3180 else 3181 err = -EPERM; 3182 break; 3183 3184 case TCP_CORK: 3185 __tcp_sock_set_cork(sk, val); 3186 break; 3187 3188 case TCP_KEEPIDLE: 3189 err = __tcp_sock_set_keepidle(sk, val); 3190 break; 3191 case TCP_KEEPINTVL: 3192 if (val < 1 || val > MAX_TCP_KEEPINTVL) 3193 err = -EINVAL; 3194 else 3195 tp->keepalive_intvl = val * HZ; 3196 break; 3197 case TCP_KEEPCNT: 3198 if (val < 1 || val > MAX_TCP_KEEPCNT) 3199 err = -EINVAL; 3200 else 3201 tp->keepalive_probes = val; 3202 break; 3203 case TCP_SYNCNT: 3204 if (val < 1 || val > MAX_TCP_SYNCNT) 3205 err = -EINVAL; 3206 else 3207 icsk->icsk_syn_retries = val; 3208 break; 3209 3210 case TCP_SAVE_SYN: 3211 if (val < 0 || val > 1) 3212 err = -EINVAL; 3213 else 3214 tp->save_syn = val; 3215 break; 3216 3217 case TCP_LINGER2: 3218 if (val < 0) 3219 tp->linger2 = -1; 3220 else if (val > TCP_FIN_TIMEOUT_MAX / HZ) 3221 tp->linger2 = TCP_FIN_TIMEOUT_MAX; 3222 else 3223 tp->linger2 = val * HZ; 3224 break; 3225 3226 case TCP_DEFER_ACCEPT: 3227 /* Translate value in seconds to number of retransmits */ 3228 icsk->icsk_accept_queue.rskq_defer_accept = 3229 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 3230 TCP_RTO_MAX / HZ); 3231 break; 3232 3233 case TCP_WINDOW_CLAMP: 3234 if (!val) { 3235 if (sk->sk_state != TCP_CLOSE) { 3236 err = -EINVAL; 3237 break; 3238 } 3239 tp->window_clamp = 0; 3240 } else 3241 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 3242 SOCK_MIN_RCVBUF / 2 : val; 3243 break; 3244 3245 case TCP_QUICKACK: 3246 __tcp_sock_set_quickack(sk, val); 3247 break; 3248 3249 #ifdef CONFIG_TCP_MD5SIG 3250 case TCP_MD5SIG: 3251 case TCP_MD5SIG_EXT: 3252 err = tp->af_specific->md5_parse(sk, optname, optval, optlen); 3253 break; 3254 #endif 3255 case TCP_USER_TIMEOUT: 3256 /* Cap the max time in ms TCP will retry or probe the window 3257 * before giving up and aborting (ETIMEDOUT) a connection. 3258 */ 3259 if (val < 0) 3260 err = -EINVAL; 3261 else 3262 icsk->icsk_user_timeout = val; 3263 break; 3264 3265 case TCP_FASTOPEN: 3266 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE | 3267 TCPF_LISTEN))) { 3268 tcp_fastopen_init_key_once(net); 3269 3270 fastopen_queue_tune(sk, val); 3271 } else { 3272 err = -EINVAL; 3273 } 3274 break; 3275 case TCP_FASTOPEN_CONNECT: 3276 if (val > 1 || val < 0) { 3277 err = -EINVAL; 3278 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) { 3279 if (sk->sk_state == TCP_CLOSE) 3280 tp->fastopen_connect = val; 3281 else 3282 err = -EINVAL; 3283 } else { 3284 err = -EOPNOTSUPP; 3285 } 3286 break; 3287 case TCP_FASTOPEN_NO_COOKIE: 3288 if (val > 1 || val < 0) 3289 err = -EINVAL; 3290 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 3291 err = -EINVAL; 3292 else 3293 tp->fastopen_no_cookie = val; 3294 break; 3295 case TCP_TIMESTAMP: 3296 if (!tp->repair) 3297 err = -EPERM; 3298 else 3299 tp->tsoffset = val - tcp_time_stamp_raw(); 3300 break; 3301 case TCP_REPAIR_WINDOW: 3302 err = tcp_repair_set_window(tp, optval, optlen); 3303 break; 3304 case TCP_NOTSENT_LOWAT: 3305 tp->notsent_lowat = val; 3306 sk->sk_write_space(sk); 3307 break; 3308 case TCP_INQ: 3309 if (val > 1 || val < 0) 3310 err = -EINVAL; 3311 else 3312 tp->recvmsg_inq = val; 3313 break; 3314 case TCP_TX_DELAY: 3315 if (val) 3316 tcp_enable_tx_delay(); 3317 tp->tcp_tx_delay = val; 3318 break; 3319 default: 3320 err = -ENOPROTOOPT; 3321 break; 3322 } 3323 3324 release_sock(sk); 3325 return err; 3326 } 3327 3328 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 3329 unsigned int optlen) 3330 { 3331 const struct inet_connection_sock *icsk = inet_csk(sk); 3332 3333 if (level != SOL_TCP) 3334 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 3335 optval, optlen); 3336 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 3337 } 3338 EXPORT_SYMBOL(tcp_setsockopt); 3339 3340 #ifdef CONFIG_COMPAT 3341 int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 3342 char __user *optval, unsigned int optlen) 3343 { 3344 if (level != SOL_TCP) 3345 return inet_csk_compat_setsockopt(sk, level, optname, 3346 optval, optlen); 3347 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 3348 } 3349 EXPORT_SYMBOL(compat_tcp_setsockopt); 3350 #endif 3351 3352 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp, 3353 struct tcp_info *info) 3354 { 3355 u64 stats[__TCP_CHRONO_MAX], total = 0; 3356 enum tcp_chrono i; 3357 3358 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) { 3359 stats[i] = tp->chrono_stat[i - 1]; 3360 if (i == tp->chrono_type) 3361 stats[i] += tcp_jiffies32 - tp->chrono_start; 3362 stats[i] *= USEC_PER_SEC / HZ; 3363 total += stats[i]; 3364 } 3365 3366 info->tcpi_busy_time = total; 3367 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED]; 3368 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED]; 3369 } 3370 3371 /* Return information about state of tcp endpoint in API format. */ 3372 void tcp_get_info(struct sock *sk, struct tcp_info *info) 3373 { 3374 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */ 3375 const struct inet_connection_sock *icsk = inet_csk(sk); 3376 unsigned long rate; 3377 u32 now; 3378 u64 rate64; 3379 bool slow; 3380 3381 memset(info, 0, sizeof(*info)); 3382 if (sk->sk_type != SOCK_STREAM) 3383 return; 3384 3385 info->tcpi_state = inet_sk_state_load(sk); 3386 3387 /* Report meaningful fields for all TCP states, including listeners */ 3388 rate = READ_ONCE(sk->sk_pacing_rate); 3389 rate64 = (rate != ~0UL) ? rate : ~0ULL; 3390 info->tcpi_pacing_rate = rate64; 3391 3392 rate = READ_ONCE(sk->sk_max_pacing_rate); 3393 rate64 = (rate != ~0UL) ? rate : ~0ULL; 3394 info->tcpi_max_pacing_rate = rate64; 3395 3396 info->tcpi_reordering = tp->reordering; 3397 info->tcpi_snd_cwnd = tp->snd_cwnd; 3398 3399 if (info->tcpi_state == TCP_LISTEN) { 3400 /* listeners aliased fields : 3401 * tcpi_unacked -> Number of children ready for accept() 3402 * tcpi_sacked -> max backlog 3403 */ 3404 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog); 3405 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog); 3406 return; 3407 } 3408 3409 slow = lock_sock_fast(sk); 3410 3411 info->tcpi_ca_state = icsk->icsk_ca_state; 3412 info->tcpi_retransmits = icsk->icsk_retransmits; 3413 info->tcpi_probes = icsk->icsk_probes_out; 3414 info->tcpi_backoff = icsk->icsk_backoff; 3415 3416 if (tp->rx_opt.tstamp_ok) 3417 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 3418 if (tcp_is_sack(tp)) 3419 info->tcpi_options |= TCPI_OPT_SACK; 3420 if (tp->rx_opt.wscale_ok) { 3421 info->tcpi_options |= TCPI_OPT_WSCALE; 3422 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 3423 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 3424 } 3425 3426 if (tp->ecn_flags & TCP_ECN_OK) 3427 info->tcpi_options |= TCPI_OPT_ECN; 3428 if (tp->ecn_flags & TCP_ECN_SEEN) 3429 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 3430 if (tp->syn_data_acked) 3431 info->tcpi_options |= TCPI_OPT_SYN_DATA; 3432 3433 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 3434 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 3435 info->tcpi_snd_mss = tp->mss_cache; 3436 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 3437 3438 info->tcpi_unacked = tp->packets_out; 3439 info->tcpi_sacked = tp->sacked_out; 3440 3441 info->tcpi_lost = tp->lost_out; 3442 info->tcpi_retrans = tp->retrans_out; 3443 3444 now = tcp_jiffies32; 3445 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 3446 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 3447 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 3448 3449 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 3450 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 3451 info->tcpi_rtt = tp->srtt_us >> 3; 3452 info->tcpi_rttvar = tp->mdev_us >> 2; 3453 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 3454 info->tcpi_advmss = tp->advmss; 3455 3456 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3; 3457 info->tcpi_rcv_space = tp->rcvq_space.space; 3458 3459 info->tcpi_total_retrans = tp->total_retrans; 3460 3461 info->tcpi_bytes_acked = tp->bytes_acked; 3462 info->tcpi_bytes_received = tp->bytes_received; 3463 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt); 3464 tcp_get_info_chrono_stats(tp, info); 3465 3466 info->tcpi_segs_out = tp->segs_out; 3467 info->tcpi_segs_in = tp->segs_in; 3468 3469 info->tcpi_min_rtt = tcp_min_rtt(tp); 3470 info->tcpi_data_segs_in = tp->data_segs_in; 3471 info->tcpi_data_segs_out = tp->data_segs_out; 3472 3473 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0; 3474 rate64 = tcp_compute_delivery_rate(tp); 3475 if (rate64) 3476 info->tcpi_delivery_rate = rate64; 3477 info->tcpi_delivered = tp->delivered; 3478 info->tcpi_delivered_ce = tp->delivered_ce; 3479 info->tcpi_bytes_sent = tp->bytes_sent; 3480 info->tcpi_bytes_retrans = tp->bytes_retrans; 3481 info->tcpi_dsack_dups = tp->dsack_dups; 3482 info->tcpi_reord_seen = tp->reord_seen; 3483 info->tcpi_rcv_ooopack = tp->rcv_ooopack; 3484 info->tcpi_snd_wnd = tp->snd_wnd; 3485 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail; 3486 unlock_sock_fast(sk, slow); 3487 } 3488 EXPORT_SYMBOL_GPL(tcp_get_info); 3489 3490 static size_t tcp_opt_stats_get_size(void) 3491 { 3492 return 3493 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */ 3494 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */ 3495 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */ 3496 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */ 3497 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */ 3498 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */ 3499 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */ 3500 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */ 3501 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */ 3502 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */ 3503 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */ 3504 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */ 3505 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */ 3506 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */ 3507 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */ 3508 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */ 3509 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */ 3510 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */ 3511 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */ 3512 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */ 3513 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */ 3514 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */ 3515 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */ 3516 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */ 3517 0; 3518 } 3519 3520 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk) 3521 { 3522 const struct tcp_sock *tp = tcp_sk(sk); 3523 struct sk_buff *stats; 3524 struct tcp_info info; 3525 unsigned long rate; 3526 u64 rate64; 3527 3528 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC); 3529 if (!stats) 3530 return NULL; 3531 3532 tcp_get_info_chrono_stats(tp, &info); 3533 nla_put_u64_64bit(stats, TCP_NLA_BUSY, 3534 info.tcpi_busy_time, TCP_NLA_PAD); 3535 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED, 3536 info.tcpi_rwnd_limited, TCP_NLA_PAD); 3537 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED, 3538 info.tcpi_sndbuf_limited, TCP_NLA_PAD); 3539 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT, 3540 tp->data_segs_out, TCP_NLA_PAD); 3541 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS, 3542 tp->total_retrans, TCP_NLA_PAD); 3543 3544 rate = READ_ONCE(sk->sk_pacing_rate); 3545 rate64 = (rate != ~0UL) ? rate : ~0ULL; 3546 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD); 3547 3548 rate64 = tcp_compute_delivery_rate(tp); 3549 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD); 3550 3551 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd); 3552 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering); 3553 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp)); 3554 3555 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits); 3556 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited); 3557 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh); 3558 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered); 3559 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce); 3560 3561 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una); 3562 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state); 3563 3564 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent, 3565 TCP_NLA_PAD); 3566 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans, 3567 TCP_NLA_PAD); 3568 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups); 3569 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen); 3570 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3); 3571 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash); 3572 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT, 3573 max_t(int, 0, tp->write_seq - tp->snd_nxt)); 3574 3575 return stats; 3576 } 3577 3578 static int do_tcp_getsockopt(struct sock *sk, int level, 3579 int optname, char __user *optval, int __user *optlen) 3580 { 3581 struct inet_connection_sock *icsk = inet_csk(sk); 3582 struct tcp_sock *tp = tcp_sk(sk); 3583 struct net *net = sock_net(sk); 3584 int val, len; 3585 3586 if (get_user(len, optlen)) 3587 return -EFAULT; 3588 3589 len = min_t(unsigned int, len, sizeof(int)); 3590 3591 if (len < 0) 3592 return -EINVAL; 3593 3594 switch (optname) { 3595 case TCP_MAXSEG: 3596 val = tp->mss_cache; 3597 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 3598 val = tp->rx_opt.user_mss; 3599 if (tp->repair) 3600 val = tp->rx_opt.mss_clamp; 3601 break; 3602 case TCP_NODELAY: 3603 val = !!(tp->nonagle&TCP_NAGLE_OFF); 3604 break; 3605 case TCP_CORK: 3606 val = !!(tp->nonagle&TCP_NAGLE_CORK); 3607 break; 3608 case TCP_KEEPIDLE: 3609 val = keepalive_time_when(tp) / HZ; 3610 break; 3611 case TCP_KEEPINTVL: 3612 val = keepalive_intvl_when(tp) / HZ; 3613 break; 3614 case TCP_KEEPCNT: 3615 val = keepalive_probes(tp); 3616 break; 3617 case TCP_SYNCNT: 3618 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries; 3619 break; 3620 case TCP_LINGER2: 3621 val = tp->linger2; 3622 if (val >= 0) 3623 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ; 3624 break; 3625 case TCP_DEFER_ACCEPT: 3626 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 3627 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 3628 break; 3629 case TCP_WINDOW_CLAMP: 3630 val = tp->window_clamp; 3631 break; 3632 case TCP_INFO: { 3633 struct tcp_info info; 3634 3635 if (get_user(len, optlen)) 3636 return -EFAULT; 3637 3638 tcp_get_info(sk, &info); 3639 3640 len = min_t(unsigned int, len, sizeof(info)); 3641 if (put_user(len, optlen)) 3642 return -EFAULT; 3643 if (copy_to_user(optval, &info, len)) 3644 return -EFAULT; 3645 return 0; 3646 } 3647 case TCP_CC_INFO: { 3648 const struct tcp_congestion_ops *ca_ops; 3649 union tcp_cc_info info; 3650 size_t sz = 0; 3651 int attr; 3652 3653 if (get_user(len, optlen)) 3654 return -EFAULT; 3655 3656 ca_ops = icsk->icsk_ca_ops; 3657 if (ca_ops && ca_ops->get_info) 3658 sz = ca_ops->get_info(sk, ~0U, &attr, &info); 3659 3660 len = min_t(unsigned int, len, sz); 3661 if (put_user(len, optlen)) 3662 return -EFAULT; 3663 if (copy_to_user(optval, &info, len)) 3664 return -EFAULT; 3665 return 0; 3666 } 3667 case TCP_QUICKACK: 3668 val = !inet_csk_in_pingpong_mode(sk); 3669 break; 3670 3671 case TCP_CONGESTION: 3672 if (get_user(len, optlen)) 3673 return -EFAULT; 3674 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 3675 if (put_user(len, optlen)) 3676 return -EFAULT; 3677 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 3678 return -EFAULT; 3679 return 0; 3680 3681 case TCP_ULP: 3682 if (get_user(len, optlen)) 3683 return -EFAULT; 3684 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX); 3685 if (!icsk->icsk_ulp_ops) { 3686 if (put_user(0, optlen)) 3687 return -EFAULT; 3688 return 0; 3689 } 3690 if (put_user(len, optlen)) 3691 return -EFAULT; 3692 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len)) 3693 return -EFAULT; 3694 return 0; 3695 3696 case TCP_FASTOPEN_KEY: { 3697 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH]; 3698 struct tcp_fastopen_context *ctx; 3699 unsigned int key_len = 0; 3700 3701 if (get_user(len, optlen)) 3702 return -EFAULT; 3703 3704 rcu_read_lock(); 3705 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx); 3706 if (ctx) { 3707 key_len = tcp_fastopen_context_len(ctx) * 3708 TCP_FASTOPEN_KEY_LENGTH; 3709 memcpy(&key[0], &ctx->key[0], key_len); 3710 } 3711 rcu_read_unlock(); 3712 3713 len = min_t(unsigned int, len, key_len); 3714 if (put_user(len, optlen)) 3715 return -EFAULT; 3716 if (copy_to_user(optval, key, len)) 3717 return -EFAULT; 3718 return 0; 3719 } 3720 case TCP_THIN_LINEAR_TIMEOUTS: 3721 val = tp->thin_lto; 3722 break; 3723 3724 case TCP_THIN_DUPACK: 3725 val = 0; 3726 break; 3727 3728 case TCP_REPAIR: 3729 val = tp->repair; 3730 break; 3731 3732 case TCP_REPAIR_QUEUE: 3733 if (tp->repair) 3734 val = tp->repair_queue; 3735 else 3736 return -EINVAL; 3737 break; 3738 3739 case TCP_REPAIR_WINDOW: { 3740 struct tcp_repair_window opt; 3741 3742 if (get_user(len, optlen)) 3743 return -EFAULT; 3744 3745 if (len != sizeof(opt)) 3746 return -EINVAL; 3747 3748 if (!tp->repair) 3749 return -EPERM; 3750 3751 opt.snd_wl1 = tp->snd_wl1; 3752 opt.snd_wnd = tp->snd_wnd; 3753 opt.max_window = tp->max_window; 3754 opt.rcv_wnd = tp->rcv_wnd; 3755 opt.rcv_wup = tp->rcv_wup; 3756 3757 if (copy_to_user(optval, &opt, len)) 3758 return -EFAULT; 3759 return 0; 3760 } 3761 case TCP_QUEUE_SEQ: 3762 if (tp->repair_queue == TCP_SEND_QUEUE) 3763 val = tp->write_seq; 3764 else if (tp->repair_queue == TCP_RECV_QUEUE) 3765 val = tp->rcv_nxt; 3766 else 3767 return -EINVAL; 3768 break; 3769 3770 case TCP_USER_TIMEOUT: 3771 val = icsk->icsk_user_timeout; 3772 break; 3773 3774 case TCP_FASTOPEN: 3775 val = icsk->icsk_accept_queue.fastopenq.max_qlen; 3776 break; 3777 3778 case TCP_FASTOPEN_CONNECT: 3779 val = tp->fastopen_connect; 3780 break; 3781 3782 case TCP_FASTOPEN_NO_COOKIE: 3783 val = tp->fastopen_no_cookie; 3784 break; 3785 3786 case TCP_TX_DELAY: 3787 val = tp->tcp_tx_delay; 3788 break; 3789 3790 case TCP_TIMESTAMP: 3791 val = tcp_time_stamp_raw() + tp->tsoffset; 3792 break; 3793 case TCP_NOTSENT_LOWAT: 3794 val = tp->notsent_lowat; 3795 break; 3796 case TCP_INQ: 3797 val = tp->recvmsg_inq; 3798 break; 3799 case TCP_SAVE_SYN: 3800 val = tp->save_syn; 3801 break; 3802 case TCP_SAVED_SYN: { 3803 if (get_user(len, optlen)) 3804 return -EFAULT; 3805 3806 lock_sock(sk); 3807 if (tp->saved_syn) { 3808 if (len < tp->saved_syn[0]) { 3809 if (put_user(tp->saved_syn[0], optlen)) { 3810 release_sock(sk); 3811 return -EFAULT; 3812 } 3813 release_sock(sk); 3814 return -EINVAL; 3815 } 3816 len = tp->saved_syn[0]; 3817 if (put_user(len, optlen)) { 3818 release_sock(sk); 3819 return -EFAULT; 3820 } 3821 if (copy_to_user(optval, tp->saved_syn + 1, len)) { 3822 release_sock(sk); 3823 return -EFAULT; 3824 } 3825 tcp_saved_syn_free(tp); 3826 release_sock(sk); 3827 } else { 3828 release_sock(sk); 3829 len = 0; 3830 if (put_user(len, optlen)) 3831 return -EFAULT; 3832 } 3833 return 0; 3834 } 3835 #ifdef CONFIG_MMU 3836 case TCP_ZEROCOPY_RECEIVE: { 3837 struct tcp_zerocopy_receive zc; 3838 int err; 3839 3840 if (get_user(len, optlen)) 3841 return -EFAULT; 3842 if (len < offsetofend(struct tcp_zerocopy_receive, length)) 3843 return -EINVAL; 3844 if (len > sizeof(zc)) { 3845 len = sizeof(zc); 3846 if (put_user(len, optlen)) 3847 return -EFAULT; 3848 } 3849 if (copy_from_user(&zc, optval, len)) 3850 return -EFAULT; 3851 lock_sock(sk); 3852 err = tcp_zerocopy_receive(sk, &zc); 3853 release_sock(sk); 3854 if (len == sizeof(zc)) 3855 goto zerocopy_rcv_sk_err; 3856 switch (len) { 3857 case offsetofend(struct tcp_zerocopy_receive, err): 3858 goto zerocopy_rcv_sk_err; 3859 case offsetofend(struct tcp_zerocopy_receive, inq): 3860 goto zerocopy_rcv_inq; 3861 case offsetofend(struct tcp_zerocopy_receive, length): 3862 default: 3863 goto zerocopy_rcv_out; 3864 } 3865 zerocopy_rcv_sk_err: 3866 if (!err) 3867 zc.err = sock_error(sk); 3868 zerocopy_rcv_inq: 3869 zc.inq = tcp_inq_hint(sk); 3870 zerocopy_rcv_out: 3871 if (!err && copy_to_user(optval, &zc, len)) 3872 err = -EFAULT; 3873 return err; 3874 } 3875 #endif 3876 default: 3877 return -ENOPROTOOPT; 3878 } 3879 3880 if (put_user(len, optlen)) 3881 return -EFAULT; 3882 if (copy_to_user(optval, &val, len)) 3883 return -EFAULT; 3884 return 0; 3885 } 3886 3887 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 3888 int __user *optlen) 3889 { 3890 struct inet_connection_sock *icsk = inet_csk(sk); 3891 3892 if (level != SOL_TCP) 3893 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 3894 optval, optlen); 3895 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 3896 } 3897 EXPORT_SYMBOL(tcp_getsockopt); 3898 3899 #ifdef CONFIG_COMPAT 3900 int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 3901 char __user *optval, int __user *optlen) 3902 { 3903 if (level != SOL_TCP) 3904 return inet_csk_compat_getsockopt(sk, level, optname, 3905 optval, optlen); 3906 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 3907 } 3908 EXPORT_SYMBOL(compat_tcp_getsockopt); 3909 #endif 3910 3911 #ifdef CONFIG_TCP_MD5SIG 3912 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool); 3913 static DEFINE_MUTEX(tcp_md5sig_mutex); 3914 static bool tcp_md5sig_pool_populated = false; 3915 3916 static void __tcp_alloc_md5sig_pool(void) 3917 { 3918 struct crypto_ahash *hash; 3919 int cpu; 3920 3921 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC); 3922 if (IS_ERR(hash)) 3923 return; 3924 3925 for_each_possible_cpu(cpu) { 3926 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch; 3927 struct ahash_request *req; 3928 3929 if (!scratch) { 3930 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) + 3931 sizeof(struct tcphdr), 3932 GFP_KERNEL, 3933 cpu_to_node(cpu)); 3934 if (!scratch) 3935 return; 3936 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch; 3937 } 3938 if (per_cpu(tcp_md5sig_pool, cpu).md5_req) 3939 continue; 3940 3941 req = ahash_request_alloc(hash, GFP_KERNEL); 3942 if (!req) 3943 return; 3944 3945 ahash_request_set_callback(req, 0, NULL, NULL); 3946 3947 per_cpu(tcp_md5sig_pool, cpu).md5_req = req; 3948 } 3949 /* before setting tcp_md5sig_pool_populated, we must commit all writes 3950 * to memory. See smp_rmb() in tcp_get_md5sig_pool() 3951 */ 3952 smp_wmb(); 3953 tcp_md5sig_pool_populated = true; 3954 } 3955 3956 bool tcp_alloc_md5sig_pool(void) 3957 { 3958 if (unlikely(!tcp_md5sig_pool_populated)) { 3959 mutex_lock(&tcp_md5sig_mutex); 3960 3961 if (!tcp_md5sig_pool_populated) { 3962 __tcp_alloc_md5sig_pool(); 3963 if (tcp_md5sig_pool_populated) 3964 static_branch_inc(&tcp_md5_needed); 3965 } 3966 3967 mutex_unlock(&tcp_md5sig_mutex); 3968 } 3969 return tcp_md5sig_pool_populated; 3970 } 3971 EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 3972 3973 3974 /** 3975 * tcp_get_md5sig_pool - get md5sig_pool for this user 3976 * 3977 * We use percpu structure, so if we succeed, we exit with preemption 3978 * and BH disabled, to make sure another thread or softirq handling 3979 * wont try to get same context. 3980 */ 3981 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 3982 { 3983 local_bh_disable(); 3984 3985 if (tcp_md5sig_pool_populated) { 3986 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */ 3987 smp_rmb(); 3988 return this_cpu_ptr(&tcp_md5sig_pool); 3989 } 3990 local_bh_enable(); 3991 return NULL; 3992 } 3993 EXPORT_SYMBOL(tcp_get_md5sig_pool); 3994 3995 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3996 const struct sk_buff *skb, unsigned int header_len) 3997 { 3998 struct scatterlist sg; 3999 const struct tcphdr *tp = tcp_hdr(skb); 4000 struct ahash_request *req = hp->md5_req; 4001 unsigned int i; 4002 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 4003 skb_headlen(skb) - header_len : 0; 4004 const struct skb_shared_info *shi = skb_shinfo(skb); 4005 struct sk_buff *frag_iter; 4006 4007 sg_init_table(&sg, 1); 4008 4009 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 4010 ahash_request_set_crypt(req, &sg, NULL, head_data_len); 4011 if (crypto_ahash_update(req)) 4012 return 1; 4013 4014 for (i = 0; i < shi->nr_frags; ++i) { 4015 const skb_frag_t *f = &shi->frags[i]; 4016 unsigned int offset = skb_frag_off(f); 4017 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT); 4018 4019 sg_set_page(&sg, page, skb_frag_size(f), 4020 offset_in_page(offset)); 4021 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f)); 4022 if (crypto_ahash_update(req)) 4023 return 1; 4024 } 4025 4026 skb_walk_frags(skb, frag_iter) 4027 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 4028 return 1; 4029 4030 return 0; 4031 } 4032 EXPORT_SYMBOL(tcp_md5_hash_skb_data); 4033 4034 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key) 4035 { 4036 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */ 4037 struct scatterlist sg; 4038 4039 sg_init_one(&sg, key->key, keylen); 4040 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen); 4041 4042 /* We use data_race() because tcp_md5_do_add() might change key->key under us */ 4043 return data_race(crypto_ahash_update(hp->md5_req)); 4044 } 4045 EXPORT_SYMBOL(tcp_md5_hash_key); 4046 4047 #endif 4048 4049 void tcp_done(struct sock *sk) 4050 { 4051 struct request_sock *req; 4052 4053 /* We might be called with a new socket, after 4054 * inet_csk_prepare_forced_close() has been called 4055 * so we can not use lockdep_sock_is_held(sk) 4056 */ 4057 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1); 4058 4059 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 4060 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 4061 4062 tcp_set_state(sk, TCP_CLOSE); 4063 tcp_clear_xmit_timers(sk); 4064 if (req) 4065 reqsk_fastopen_remove(sk, req, false); 4066 4067 sk->sk_shutdown = SHUTDOWN_MASK; 4068 4069 if (!sock_flag(sk, SOCK_DEAD)) 4070 sk->sk_state_change(sk); 4071 else 4072 inet_csk_destroy_sock(sk); 4073 } 4074 EXPORT_SYMBOL_GPL(tcp_done); 4075 4076 int tcp_abort(struct sock *sk, int err) 4077 { 4078 if (!sk_fullsock(sk)) { 4079 if (sk->sk_state == TCP_NEW_SYN_RECV) { 4080 struct request_sock *req = inet_reqsk(sk); 4081 4082 local_bh_disable(); 4083 inet_csk_reqsk_queue_drop(req->rsk_listener, req); 4084 local_bh_enable(); 4085 return 0; 4086 } 4087 return -EOPNOTSUPP; 4088 } 4089 4090 /* Don't race with userspace socket closes such as tcp_close. */ 4091 lock_sock(sk); 4092 4093 if (sk->sk_state == TCP_LISTEN) { 4094 tcp_set_state(sk, TCP_CLOSE); 4095 inet_csk_listen_stop(sk); 4096 } 4097 4098 /* Don't race with BH socket closes such as inet_csk_listen_stop. */ 4099 local_bh_disable(); 4100 bh_lock_sock(sk); 4101 4102 if (!sock_flag(sk, SOCK_DEAD)) { 4103 sk->sk_err = err; 4104 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4105 smp_wmb(); 4106 sk->sk_error_report(sk); 4107 if (tcp_need_reset(sk->sk_state)) 4108 tcp_send_active_reset(sk, GFP_ATOMIC); 4109 tcp_done(sk); 4110 } 4111 4112 bh_unlock_sock(sk); 4113 local_bh_enable(); 4114 tcp_write_queue_purge(sk); 4115 release_sock(sk); 4116 return 0; 4117 } 4118 EXPORT_SYMBOL_GPL(tcp_abort); 4119 4120 extern struct tcp_congestion_ops tcp_reno; 4121 4122 static __initdata unsigned long thash_entries; 4123 static int __init set_thash_entries(char *str) 4124 { 4125 ssize_t ret; 4126 4127 if (!str) 4128 return 0; 4129 4130 ret = kstrtoul(str, 0, &thash_entries); 4131 if (ret) 4132 return 0; 4133 4134 return 1; 4135 } 4136 __setup("thash_entries=", set_thash_entries); 4137 4138 static void __init tcp_init_mem(void) 4139 { 4140 unsigned long limit = nr_free_buffer_pages() / 16; 4141 4142 limit = max(limit, 128UL); 4143 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */ 4144 sysctl_tcp_mem[1] = limit; /* 6.25 % */ 4145 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */ 4146 } 4147 4148 void __init tcp_init(void) 4149 { 4150 int max_rshare, max_wshare, cnt; 4151 unsigned long limit; 4152 unsigned int i; 4153 4154 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE); 4155 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > 4156 sizeof_field(struct sk_buff, cb)); 4157 4158 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL); 4159 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL); 4160 inet_hashinfo_init(&tcp_hashinfo); 4161 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash", 4162 thash_entries, 21, /* one slot per 2 MB*/ 4163 0, 64 * 1024); 4164 tcp_hashinfo.bind_bucket_cachep = 4165 kmem_cache_create("tcp_bind_bucket", 4166 sizeof(struct inet_bind_bucket), 0, 4167 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 4168 4169 /* Size and allocate the main established and bind bucket 4170 * hash tables. 4171 * 4172 * The methodology is similar to that of the buffer cache. 4173 */ 4174 tcp_hashinfo.ehash = 4175 alloc_large_system_hash("TCP established", 4176 sizeof(struct inet_ehash_bucket), 4177 thash_entries, 4178 17, /* one slot per 128 KB of memory */ 4179 0, 4180 NULL, 4181 &tcp_hashinfo.ehash_mask, 4182 0, 4183 thash_entries ? 0 : 512 * 1024); 4184 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) 4185 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 4186 4187 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 4188 panic("TCP: failed to alloc ehash_locks"); 4189 tcp_hashinfo.bhash = 4190 alloc_large_system_hash("TCP bind", 4191 sizeof(struct inet_bind_hashbucket), 4192 tcp_hashinfo.ehash_mask + 1, 4193 17, /* one slot per 128 KB of memory */ 4194 0, 4195 &tcp_hashinfo.bhash_size, 4196 NULL, 4197 0, 4198 64 * 1024); 4199 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 4200 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 4201 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 4202 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 4203 } 4204 4205 4206 cnt = tcp_hashinfo.ehash_mask + 1; 4207 sysctl_tcp_max_orphans = cnt / 2; 4208 4209 tcp_init_mem(); 4210 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 4211 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 4212 max_wshare = min(4UL*1024*1024, limit); 4213 max_rshare = min(6UL*1024*1024, limit); 4214 4215 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 4216 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024; 4217 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare); 4218 4219 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 4220 init_net.ipv4.sysctl_tcp_rmem[1] = 131072; 4221 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare); 4222 4223 pr_info("Hash tables configured (established %u bind %u)\n", 4224 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 4225 4226 tcp_v4_init(); 4227 tcp_metrics_init(); 4228 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0); 4229 tcp_tasklet_init(); 4230 mptcp_init(); 4231 } 4232