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 down_read(mmap_sem) */ 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_receive(struct sock *sk, 1746 struct tcp_zerocopy_receive *zc) 1747 { 1748 unsigned long address = (unsigned long)zc->address; 1749 u32 length = 0, seq, offset, zap_len; 1750 const skb_frag_t *frags = NULL; 1751 struct vm_area_struct *vma; 1752 struct sk_buff *skb = NULL; 1753 struct tcp_sock *tp; 1754 int inq; 1755 int ret; 1756 1757 if (address & (PAGE_SIZE - 1) || address != zc->address) 1758 return -EINVAL; 1759 1760 if (sk->sk_state == TCP_LISTEN) 1761 return -ENOTCONN; 1762 1763 sock_rps_record_flow(sk); 1764 1765 down_read(¤t->mm->mmap_sem); 1766 1767 vma = find_vma(current->mm, address); 1768 if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops) { 1769 up_read(¤t->mm->mmap_sem); 1770 return -EINVAL; 1771 } 1772 zc->length = min_t(unsigned long, zc->length, vma->vm_end - address); 1773 1774 tp = tcp_sk(sk); 1775 seq = tp->copied_seq; 1776 inq = tcp_inq(sk); 1777 zc->length = min_t(u32, zc->length, inq); 1778 zap_len = zc->length & ~(PAGE_SIZE - 1); 1779 if (zap_len) { 1780 zap_page_range(vma, address, zap_len); 1781 zc->recv_skip_hint = 0; 1782 } else { 1783 zc->recv_skip_hint = zc->length; 1784 } 1785 ret = 0; 1786 while (length + PAGE_SIZE <= zc->length) { 1787 if (zc->recv_skip_hint < PAGE_SIZE) { 1788 if (skb) { 1789 if (zc->recv_skip_hint > 0) 1790 break; 1791 skb = skb->next; 1792 offset = seq - TCP_SKB_CB(skb)->seq; 1793 } else { 1794 skb = tcp_recv_skb(sk, seq, &offset); 1795 } 1796 1797 zc->recv_skip_hint = skb->len - offset; 1798 offset -= skb_headlen(skb); 1799 if ((int)offset < 0 || skb_has_frag_list(skb)) 1800 break; 1801 frags = skb_shinfo(skb)->frags; 1802 while (offset) { 1803 if (skb_frag_size(frags) > offset) 1804 goto out; 1805 offset -= skb_frag_size(frags); 1806 frags++; 1807 } 1808 } 1809 if (skb_frag_size(frags) != PAGE_SIZE || skb_frag_off(frags)) { 1810 int remaining = zc->recv_skip_hint; 1811 1812 while (remaining && (skb_frag_size(frags) != PAGE_SIZE || 1813 skb_frag_off(frags))) { 1814 remaining -= skb_frag_size(frags); 1815 frags++; 1816 } 1817 zc->recv_skip_hint -= remaining; 1818 break; 1819 } 1820 ret = vm_insert_page(vma, address + length, 1821 skb_frag_page(frags)); 1822 if (ret) 1823 break; 1824 length += PAGE_SIZE; 1825 seq += PAGE_SIZE; 1826 zc->recv_skip_hint -= PAGE_SIZE; 1827 frags++; 1828 } 1829 out: 1830 up_read(¤t->mm->mmap_sem); 1831 if (length) { 1832 WRITE_ONCE(tp->copied_seq, seq); 1833 tcp_rcv_space_adjust(sk); 1834 1835 /* Clean up data we have read: This will do ACK frames. */ 1836 tcp_recv_skb(sk, seq, &offset); 1837 tcp_cleanup_rbuf(sk, length); 1838 ret = 0; 1839 if (length == zc->length) 1840 zc->recv_skip_hint = 0; 1841 } else { 1842 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE)) 1843 ret = -EIO; 1844 } 1845 zc->length = length; 1846 return ret; 1847 } 1848 #endif 1849 1850 static void tcp_update_recv_tstamps(struct sk_buff *skb, 1851 struct scm_timestamping_internal *tss) 1852 { 1853 if (skb->tstamp) 1854 tss->ts[0] = ktime_to_timespec64(skb->tstamp); 1855 else 1856 tss->ts[0] = (struct timespec64) {0}; 1857 1858 if (skb_hwtstamps(skb)->hwtstamp) 1859 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp); 1860 else 1861 tss->ts[2] = (struct timespec64) {0}; 1862 } 1863 1864 /* Similar to __sock_recv_timestamp, but does not require an skb */ 1865 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk, 1866 struct scm_timestamping_internal *tss) 1867 { 1868 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW); 1869 bool has_timestamping = false; 1870 1871 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) { 1872 if (sock_flag(sk, SOCK_RCVTSTAMP)) { 1873 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) { 1874 if (new_tstamp) { 1875 struct __kernel_timespec kts = { 1876 .tv_sec = tss->ts[0].tv_sec, 1877 .tv_nsec = tss->ts[0].tv_nsec, 1878 }; 1879 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW, 1880 sizeof(kts), &kts); 1881 } else { 1882 struct __kernel_old_timespec ts_old = { 1883 .tv_sec = tss->ts[0].tv_sec, 1884 .tv_nsec = tss->ts[0].tv_nsec, 1885 }; 1886 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD, 1887 sizeof(ts_old), &ts_old); 1888 } 1889 } else { 1890 if (new_tstamp) { 1891 struct __kernel_sock_timeval stv = { 1892 .tv_sec = tss->ts[0].tv_sec, 1893 .tv_usec = tss->ts[0].tv_nsec / 1000, 1894 }; 1895 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW, 1896 sizeof(stv), &stv); 1897 } else { 1898 struct __kernel_old_timeval tv = { 1899 .tv_sec = tss->ts[0].tv_sec, 1900 .tv_usec = tss->ts[0].tv_nsec / 1000, 1901 }; 1902 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD, 1903 sizeof(tv), &tv); 1904 } 1905 } 1906 } 1907 1908 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE) 1909 has_timestamping = true; 1910 else 1911 tss->ts[0] = (struct timespec64) {0}; 1912 } 1913 1914 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) { 1915 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE) 1916 has_timestamping = true; 1917 else 1918 tss->ts[2] = (struct timespec64) {0}; 1919 } 1920 1921 if (has_timestamping) { 1922 tss->ts[1] = (struct timespec64) {0}; 1923 if (sock_flag(sk, SOCK_TSTAMP_NEW)) 1924 put_cmsg_scm_timestamping64(msg, tss); 1925 else 1926 put_cmsg_scm_timestamping(msg, tss); 1927 } 1928 } 1929 1930 static int tcp_inq_hint(struct sock *sk) 1931 { 1932 const struct tcp_sock *tp = tcp_sk(sk); 1933 u32 copied_seq = READ_ONCE(tp->copied_seq); 1934 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt); 1935 int inq; 1936 1937 inq = rcv_nxt - copied_seq; 1938 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) { 1939 lock_sock(sk); 1940 inq = tp->rcv_nxt - tp->copied_seq; 1941 release_sock(sk); 1942 } 1943 /* After receiving a FIN, tell the user-space to continue reading 1944 * by returning a non-zero inq. 1945 */ 1946 if (inq == 0 && sock_flag(sk, SOCK_DONE)) 1947 inq = 1; 1948 return inq; 1949 } 1950 1951 /* 1952 * This routine copies from a sock struct into the user buffer. 1953 * 1954 * Technical note: in 2.3 we work on _locked_ socket, so that 1955 * tricks with *seq access order and skb->users are not required. 1956 * Probably, code can be easily improved even more. 1957 */ 1958 1959 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock, 1960 int flags, int *addr_len) 1961 { 1962 struct tcp_sock *tp = tcp_sk(sk); 1963 int copied = 0; 1964 u32 peek_seq; 1965 u32 *seq; 1966 unsigned long used; 1967 int err, inq; 1968 int target; /* Read at least this many bytes */ 1969 long timeo; 1970 struct sk_buff *skb, *last; 1971 u32 urg_hole = 0; 1972 struct scm_timestamping_internal tss; 1973 int cmsg_flags; 1974 1975 if (unlikely(flags & MSG_ERRQUEUE)) 1976 return inet_recv_error(sk, msg, len, addr_len); 1977 1978 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) && 1979 (sk->sk_state == TCP_ESTABLISHED)) 1980 sk_busy_loop(sk, nonblock); 1981 1982 lock_sock(sk); 1983 1984 err = -ENOTCONN; 1985 if (sk->sk_state == TCP_LISTEN) 1986 goto out; 1987 1988 cmsg_flags = tp->recvmsg_inq ? 1 : 0; 1989 timeo = sock_rcvtimeo(sk, nonblock); 1990 1991 /* Urgent data needs to be handled specially. */ 1992 if (flags & MSG_OOB) 1993 goto recv_urg; 1994 1995 if (unlikely(tp->repair)) { 1996 err = -EPERM; 1997 if (!(flags & MSG_PEEK)) 1998 goto out; 1999 2000 if (tp->repair_queue == TCP_SEND_QUEUE) 2001 goto recv_sndq; 2002 2003 err = -EINVAL; 2004 if (tp->repair_queue == TCP_NO_QUEUE) 2005 goto out; 2006 2007 /* 'common' recv queue MSG_PEEK-ing */ 2008 } 2009 2010 seq = &tp->copied_seq; 2011 if (flags & MSG_PEEK) { 2012 peek_seq = tp->copied_seq; 2013 seq = &peek_seq; 2014 } 2015 2016 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2017 2018 do { 2019 u32 offset; 2020 2021 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 2022 if (tp->urg_data && tp->urg_seq == *seq) { 2023 if (copied) 2024 break; 2025 if (signal_pending(current)) { 2026 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 2027 break; 2028 } 2029 } 2030 2031 /* Next get a buffer. */ 2032 2033 last = skb_peek_tail(&sk->sk_receive_queue); 2034 skb_queue_walk(&sk->sk_receive_queue, skb) { 2035 last = skb; 2036 /* Now that we have two receive queues this 2037 * shouldn't happen. 2038 */ 2039 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 2040 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n", 2041 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 2042 flags)) 2043 break; 2044 2045 offset = *seq - TCP_SKB_CB(skb)->seq; 2046 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 2047 pr_err_once("%s: found a SYN, please report !\n", __func__); 2048 offset--; 2049 } 2050 if (offset < skb->len) 2051 goto found_ok_skb; 2052 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2053 goto found_fin_ok; 2054 WARN(!(flags & MSG_PEEK), 2055 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n", 2056 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 2057 } 2058 2059 /* Well, if we have backlog, try to process it now yet. */ 2060 2061 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail)) 2062 break; 2063 2064 if (copied) { 2065 if (sk->sk_err || 2066 sk->sk_state == TCP_CLOSE || 2067 (sk->sk_shutdown & RCV_SHUTDOWN) || 2068 !timeo || 2069 signal_pending(current)) 2070 break; 2071 } else { 2072 if (sock_flag(sk, SOCK_DONE)) 2073 break; 2074 2075 if (sk->sk_err) { 2076 copied = sock_error(sk); 2077 break; 2078 } 2079 2080 if (sk->sk_shutdown & RCV_SHUTDOWN) 2081 break; 2082 2083 if (sk->sk_state == TCP_CLOSE) { 2084 /* This occurs when user tries to read 2085 * from never connected socket. 2086 */ 2087 copied = -ENOTCONN; 2088 break; 2089 } 2090 2091 if (!timeo) { 2092 copied = -EAGAIN; 2093 break; 2094 } 2095 2096 if (signal_pending(current)) { 2097 copied = sock_intr_errno(timeo); 2098 break; 2099 } 2100 } 2101 2102 tcp_cleanup_rbuf(sk, copied); 2103 2104 if (copied >= target) { 2105 /* Do not sleep, just process backlog. */ 2106 release_sock(sk); 2107 lock_sock(sk); 2108 } else { 2109 sk_wait_data(sk, &timeo, last); 2110 } 2111 2112 if ((flags & MSG_PEEK) && 2113 (peek_seq - copied - urg_hole != tp->copied_seq)) { 2114 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n", 2115 current->comm, 2116 task_pid_nr(current)); 2117 peek_seq = tp->copied_seq; 2118 } 2119 continue; 2120 2121 found_ok_skb: 2122 /* Ok so how much can we use? */ 2123 used = skb->len - offset; 2124 if (len < used) 2125 used = len; 2126 2127 /* Do we have urgent data here? */ 2128 if (tp->urg_data) { 2129 u32 urg_offset = tp->urg_seq - *seq; 2130 if (urg_offset < used) { 2131 if (!urg_offset) { 2132 if (!sock_flag(sk, SOCK_URGINLINE)) { 2133 WRITE_ONCE(*seq, *seq + 1); 2134 urg_hole++; 2135 offset++; 2136 used--; 2137 if (!used) 2138 goto skip_copy; 2139 } 2140 } else 2141 used = urg_offset; 2142 } 2143 } 2144 2145 if (!(flags & MSG_TRUNC)) { 2146 err = skb_copy_datagram_msg(skb, offset, msg, used); 2147 if (err) { 2148 /* Exception. Bailout! */ 2149 if (!copied) 2150 copied = -EFAULT; 2151 break; 2152 } 2153 } 2154 2155 WRITE_ONCE(*seq, *seq + used); 2156 copied += used; 2157 len -= used; 2158 2159 tcp_rcv_space_adjust(sk); 2160 2161 skip_copy: 2162 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) { 2163 tp->urg_data = 0; 2164 tcp_fast_path_check(sk); 2165 } 2166 2167 if (TCP_SKB_CB(skb)->has_rxtstamp) { 2168 tcp_update_recv_tstamps(skb, &tss); 2169 cmsg_flags |= 2; 2170 } 2171 2172 if (used + offset < skb->len) 2173 continue; 2174 2175 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2176 goto found_fin_ok; 2177 if (!(flags & MSG_PEEK)) 2178 sk_eat_skb(sk, skb); 2179 continue; 2180 2181 found_fin_ok: 2182 /* Process the FIN. */ 2183 WRITE_ONCE(*seq, *seq + 1); 2184 if (!(flags & MSG_PEEK)) 2185 sk_eat_skb(sk, skb); 2186 break; 2187 } while (len > 0); 2188 2189 /* According to UNIX98, msg_name/msg_namelen are ignored 2190 * on connected socket. I was just happy when found this 8) --ANK 2191 */ 2192 2193 /* Clean up data we have read: This will do ACK frames. */ 2194 tcp_cleanup_rbuf(sk, copied); 2195 2196 release_sock(sk); 2197 2198 if (cmsg_flags) { 2199 if (cmsg_flags & 2) 2200 tcp_recv_timestamp(msg, sk, &tss); 2201 if (cmsg_flags & 1) { 2202 inq = tcp_inq_hint(sk); 2203 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq); 2204 } 2205 } 2206 2207 return copied; 2208 2209 out: 2210 release_sock(sk); 2211 return err; 2212 2213 recv_urg: 2214 err = tcp_recv_urg(sk, msg, len, flags); 2215 goto out; 2216 2217 recv_sndq: 2218 err = tcp_peek_sndq(sk, msg, len); 2219 goto out; 2220 } 2221 EXPORT_SYMBOL(tcp_recvmsg); 2222 2223 void tcp_set_state(struct sock *sk, int state) 2224 { 2225 int oldstate = sk->sk_state; 2226 2227 /* We defined a new enum for TCP states that are exported in BPF 2228 * so as not force the internal TCP states to be frozen. The 2229 * following checks will detect if an internal state value ever 2230 * differs from the BPF value. If this ever happens, then we will 2231 * need to remap the internal value to the BPF value before calling 2232 * tcp_call_bpf_2arg. 2233 */ 2234 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED); 2235 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT); 2236 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV); 2237 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1); 2238 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2); 2239 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT); 2240 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE); 2241 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT); 2242 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK); 2243 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN); 2244 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING); 2245 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV); 2246 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES); 2247 2248 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG)) 2249 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state); 2250 2251 switch (state) { 2252 case TCP_ESTABLISHED: 2253 if (oldstate != TCP_ESTABLISHED) 2254 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 2255 break; 2256 2257 case TCP_CLOSE: 2258 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 2259 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 2260 2261 sk->sk_prot->unhash(sk); 2262 if (inet_csk(sk)->icsk_bind_hash && 2263 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 2264 inet_put_port(sk); 2265 fallthrough; 2266 default: 2267 if (oldstate == TCP_ESTABLISHED) 2268 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 2269 } 2270 2271 /* Change state AFTER socket is unhashed to avoid closed 2272 * socket sitting in hash tables. 2273 */ 2274 inet_sk_state_store(sk, state); 2275 } 2276 EXPORT_SYMBOL_GPL(tcp_set_state); 2277 2278 /* 2279 * State processing on a close. This implements the state shift for 2280 * sending our FIN frame. Note that we only send a FIN for some 2281 * states. A shutdown() may have already sent the FIN, or we may be 2282 * closed. 2283 */ 2284 2285 static const unsigned char new_state[16] = { 2286 /* current state: new state: action: */ 2287 [0 /* (Invalid) */] = TCP_CLOSE, 2288 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2289 [TCP_SYN_SENT] = TCP_CLOSE, 2290 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2291 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 2292 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 2293 [TCP_TIME_WAIT] = TCP_CLOSE, 2294 [TCP_CLOSE] = TCP_CLOSE, 2295 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 2296 [TCP_LAST_ACK] = TCP_LAST_ACK, 2297 [TCP_LISTEN] = TCP_CLOSE, 2298 [TCP_CLOSING] = TCP_CLOSING, 2299 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 2300 }; 2301 2302 static int tcp_close_state(struct sock *sk) 2303 { 2304 int next = (int)new_state[sk->sk_state]; 2305 int ns = next & TCP_STATE_MASK; 2306 2307 tcp_set_state(sk, ns); 2308 2309 return next & TCP_ACTION_FIN; 2310 } 2311 2312 /* 2313 * Shutdown the sending side of a connection. Much like close except 2314 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 2315 */ 2316 2317 void tcp_shutdown(struct sock *sk, int how) 2318 { 2319 /* We need to grab some memory, and put together a FIN, 2320 * and then put it into the queue to be sent. 2321 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 2322 */ 2323 if (!(how & SEND_SHUTDOWN)) 2324 return; 2325 2326 /* If we've already sent a FIN, or it's a closed state, skip this. */ 2327 if ((1 << sk->sk_state) & 2328 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 2329 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) { 2330 /* Clear out any half completed packets. FIN if needed. */ 2331 if (tcp_close_state(sk)) 2332 tcp_send_fin(sk); 2333 } 2334 } 2335 EXPORT_SYMBOL(tcp_shutdown); 2336 2337 bool tcp_check_oom(struct sock *sk, int shift) 2338 { 2339 bool too_many_orphans, out_of_socket_memory; 2340 2341 too_many_orphans = tcp_too_many_orphans(sk, shift); 2342 out_of_socket_memory = tcp_out_of_memory(sk); 2343 2344 if (too_many_orphans) 2345 net_info_ratelimited("too many orphaned sockets\n"); 2346 if (out_of_socket_memory) 2347 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n"); 2348 return too_many_orphans || out_of_socket_memory; 2349 } 2350 2351 void tcp_close(struct sock *sk, long timeout) 2352 { 2353 struct sk_buff *skb; 2354 int data_was_unread = 0; 2355 int state; 2356 2357 lock_sock(sk); 2358 sk->sk_shutdown = SHUTDOWN_MASK; 2359 2360 if (sk->sk_state == TCP_LISTEN) { 2361 tcp_set_state(sk, TCP_CLOSE); 2362 2363 /* Special case. */ 2364 inet_csk_listen_stop(sk); 2365 2366 goto adjudge_to_death; 2367 } 2368 2369 /* We need to flush the recv. buffs. We do this only on the 2370 * descriptor close, not protocol-sourced closes, because the 2371 * reader process may not have drained the data yet! 2372 */ 2373 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) { 2374 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq; 2375 2376 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2377 len--; 2378 data_was_unread += len; 2379 __kfree_skb(skb); 2380 } 2381 2382 sk_mem_reclaim(sk); 2383 2384 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 2385 if (sk->sk_state == TCP_CLOSE) 2386 goto adjudge_to_death; 2387 2388 /* As outlined in RFC 2525, section 2.17, we send a RST here because 2389 * data was lost. To witness the awful effects of the old behavior of 2390 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 2391 * GET in an FTP client, suspend the process, wait for the client to 2392 * advertise a zero window, then kill -9 the FTP client, wheee... 2393 * Note: timeout is always zero in such a case. 2394 */ 2395 if (unlikely(tcp_sk(sk)->repair)) { 2396 sk->sk_prot->disconnect(sk, 0); 2397 } else if (data_was_unread) { 2398 /* Unread data was tossed, zap the connection. */ 2399 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 2400 tcp_set_state(sk, TCP_CLOSE); 2401 tcp_send_active_reset(sk, sk->sk_allocation); 2402 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 2403 /* Check zero linger _after_ checking for unread data. */ 2404 sk->sk_prot->disconnect(sk, 0); 2405 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 2406 } else if (tcp_close_state(sk)) { 2407 /* We FIN if the application ate all the data before 2408 * zapping the connection. 2409 */ 2410 2411 /* RED-PEN. Formally speaking, we have broken TCP state 2412 * machine. State transitions: 2413 * 2414 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 2415 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible) 2416 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 2417 * 2418 * are legal only when FIN has been sent (i.e. in window), 2419 * rather than queued out of window. Purists blame. 2420 * 2421 * F.e. "RFC state" is ESTABLISHED, 2422 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 2423 * 2424 * The visible declinations are that sometimes 2425 * we enter time-wait state, when it is not required really 2426 * (harmless), do not send active resets, when they are 2427 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 2428 * they look as CLOSING or LAST_ACK for Linux) 2429 * Probably, I missed some more holelets. 2430 * --ANK 2431 * XXX (TFO) - To start off we don't support SYN+ACK+FIN 2432 * in a single packet! (May consider it later but will 2433 * probably need API support or TCP_CORK SYN-ACK until 2434 * data is written and socket is closed.) 2435 */ 2436 tcp_send_fin(sk); 2437 } 2438 2439 sk_stream_wait_close(sk, timeout); 2440 2441 adjudge_to_death: 2442 state = sk->sk_state; 2443 sock_hold(sk); 2444 sock_orphan(sk); 2445 2446 local_bh_disable(); 2447 bh_lock_sock(sk); 2448 /* remove backlog if any, without releasing ownership. */ 2449 __release_sock(sk); 2450 2451 percpu_counter_inc(sk->sk_prot->orphan_count); 2452 2453 /* Have we already been destroyed by a softirq or backlog? */ 2454 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 2455 goto out; 2456 2457 /* This is a (useful) BSD violating of the RFC. There is a 2458 * problem with TCP as specified in that the other end could 2459 * keep a socket open forever with no application left this end. 2460 * We use a 1 minute timeout (about the same as BSD) then kill 2461 * our end. If they send after that then tough - BUT: long enough 2462 * that we won't make the old 4*rto = almost no time - whoops 2463 * reset mistake. 2464 * 2465 * Nope, it was not mistake. It is really desired behaviour 2466 * f.e. on http servers, when such sockets are useless, but 2467 * consume significant resources. Let's do it with special 2468 * linger2 option. --ANK 2469 */ 2470 2471 if (sk->sk_state == TCP_FIN_WAIT2) { 2472 struct tcp_sock *tp = tcp_sk(sk); 2473 if (tp->linger2 < 0) { 2474 tcp_set_state(sk, TCP_CLOSE); 2475 tcp_send_active_reset(sk, GFP_ATOMIC); 2476 __NET_INC_STATS(sock_net(sk), 2477 LINUX_MIB_TCPABORTONLINGER); 2478 } else { 2479 const int tmo = tcp_fin_time(sk); 2480 2481 if (tmo > TCP_TIMEWAIT_LEN) { 2482 inet_csk_reset_keepalive_timer(sk, 2483 tmo - TCP_TIMEWAIT_LEN); 2484 } else { 2485 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 2486 goto out; 2487 } 2488 } 2489 } 2490 if (sk->sk_state != TCP_CLOSE) { 2491 sk_mem_reclaim(sk); 2492 if (tcp_check_oom(sk, 0)) { 2493 tcp_set_state(sk, TCP_CLOSE); 2494 tcp_send_active_reset(sk, GFP_ATOMIC); 2495 __NET_INC_STATS(sock_net(sk), 2496 LINUX_MIB_TCPABORTONMEMORY); 2497 } else if (!check_net(sock_net(sk))) { 2498 /* Not possible to send reset; just close */ 2499 tcp_set_state(sk, TCP_CLOSE); 2500 } 2501 } 2502 2503 if (sk->sk_state == TCP_CLOSE) { 2504 struct request_sock *req; 2505 2506 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 2507 lockdep_sock_is_held(sk)); 2508 /* We could get here with a non-NULL req if the socket is 2509 * aborted (e.g., closed with unread data) before 3WHS 2510 * finishes. 2511 */ 2512 if (req) 2513 reqsk_fastopen_remove(sk, req, false); 2514 inet_csk_destroy_sock(sk); 2515 } 2516 /* Otherwise, socket is reprieved until protocol close. */ 2517 2518 out: 2519 bh_unlock_sock(sk); 2520 local_bh_enable(); 2521 release_sock(sk); 2522 sock_put(sk); 2523 } 2524 EXPORT_SYMBOL(tcp_close); 2525 2526 /* These states need RST on ABORT according to RFC793 */ 2527 2528 static inline bool tcp_need_reset(int state) 2529 { 2530 return (1 << state) & 2531 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 2532 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 2533 } 2534 2535 static void tcp_rtx_queue_purge(struct sock *sk) 2536 { 2537 struct rb_node *p = rb_first(&sk->tcp_rtx_queue); 2538 2539 tcp_sk(sk)->highest_sack = NULL; 2540 while (p) { 2541 struct sk_buff *skb = rb_to_skb(p); 2542 2543 p = rb_next(p); 2544 /* Since we are deleting whole queue, no need to 2545 * list_del(&skb->tcp_tsorted_anchor) 2546 */ 2547 tcp_rtx_queue_unlink(skb, sk); 2548 sk_wmem_free_skb(sk, skb); 2549 } 2550 } 2551 2552 void tcp_write_queue_purge(struct sock *sk) 2553 { 2554 struct sk_buff *skb; 2555 2556 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 2557 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) { 2558 tcp_skb_tsorted_anchor_cleanup(skb); 2559 sk_wmem_free_skb(sk, skb); 2560 } 2561 tcp_rtx_queue_purge(sk); 2562 skb = sk->sk_tx_skb_cache; 2563 if (skb) { 2564 __kfree_skb(skb); 2565 sk->sk_tx_skb_cache = NULL; 2566 } 2567 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue); 2568 sk_mem_reclaim(sk); 2569 tcp_clear_all_retrans_hints(tcp_sk(sk)); 2570 tcp_sk(sk)->packets_out = 0; 2571 inet_csk(sk)->icsk_backoff = 0; 2572 } 2573 2574 int tcp_disconnect(struct sock *sk, int flags) 2575 { 2576 struct inet_sock *inet = inet_sk(sk); 2577 struct inet_connection_sock *icsk = inet_csk(sk); 2578 struct tcp_sock *tp = tcp_sk(sk); 2579 int old_state = sk->sk_state; 2580 u32 seq; 2581 2582 if (old_state != TCP_CLOSE) 2583 tcp_set_state(sk, TCP_CLOSE); 2584 2585 /* ABORT function of RFC793 */ 2586 if (old_state == TCP_LISTEN) { 2587 inet_csk_listen_stop(sk); 2588 } else if (unlikely(tp->repair)) { 2589 sk->sk_err = ECONNABORTED; 2590 } else if (tcp_need_reset(old_state) || 2591 (tp->snd_nxt != tp->write_seq && 2592 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) { 2593 /* The last check adjusts for discrepancy of Linux wrt. RFC 2594 * states 2595 */ 2596 tcp_send_active_reset(sk, gfp_any()); 2597 sk->sk_err = ECONNRESET; 2598 } else if (old_state == TCP_SYN_SENT) 2599 sk->sk_err = ECONNRESET; 2600 2601 tcp_clear_xmit_timers(sk); 2602 __skb_queue_purge(&sk->sk_receive_queue); 2603 if (sk->sk_rx_skb_cache) { 2604 __kfree_skb(sk->sk_rx_skb_cache); 2605 sk->sk_rx_skb_cache = NULL; 2606 } 2607 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 2608 tp->urg_data = 0; 2609 tcp_write_queue_purge(sk); 2610 tcp_fastopen_active_disable_ofo_check(sk); 2611 skb_rbtree_purge(&tp->out_of_order_queue); 2612 2613 inet->inet_dport = 0; 2614 2615 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) 2616 inet_reset_saddr(sk); 2617 2618 sk->sk_shutdown = 0; 2619 sock_reset_flag(sk, SOCK_DONE); 2620 tp->srtt_us = 0; 2621 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 2622 tp->rcv_rtt_last_tsecr = 0; 2623 2624 seq = tp->write_seq + tp->max_window + 2; 2625 if (!seq) 2626 seq = 1; 2627 WRITE_ONCE(tp->write_seq, seq); 2628 2629 icsk->icsk_backoff = 0; 2630 icsk->icsk_probes_out = 0; 2631 icsk->icsk_rto = TCP_TIMEOUT_INIT; 2632 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 2633 tp->snd_cwnd = TCP_INIT_CWND; 2634 tp->snd_cwnd_cnt = 0; 2635 tp->window_clamp = 0; 2636 tp->delivered = 0; 2637 tp->delivered_ce = 0; 2638 tcp_set_ca_state(sk, TCP_CA_Open); 2639 tp->is_sack_reneg = 0; 2640 tcp_clear_retrans(tp); 2641 tp->total_retrans = 0; 2642 inet_csk_delack_init(sk); 2643 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0 2644 * issue in __tcp_select_window() 2645 */ 2646 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS; 2647 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 2648 __sk_dst_reset(sk); 2649 dst_release(sk->sk_rx_dst); 2650 sk->sk_rx_dst = NULL; 2651 tcp_saved_syn_free(tp); 2652 tp->compressed_ack = 0; 2653 tp->segs_in = 0; 2654 tp->segs_out = 0; 2655 tp->bytes_sent = 0; 2656 tp->bytes_acked = 0; 2657 tp->bytes_received = 0; 2658 tp->bytes_retrans = 0; 2659 tp->data_segs_in = 0; 2660 tp->data_segs_out = 0; 2661 tp->duplicate_sack[0].start_seq = 0; 2662 tp->duplicate_sack[0].end_seq = 0; 2663 tp->dsack_dups = 0; 2664 tp->reord_seen = 0; 2665 tp->retrans_out = 0; 2666 tp->sacked_out = 0; 2667 tp->tlp_high_seq = 0; 2668 tp->last_oow_ack_time = 0; 2669 /* There's a bubble in the pipe until at least the first ACK. */ 2670 tp->app_limited = ~0U; 2671 tp->rack.mstamp = 0; 2672 tp->rack.advanced = 0; 2673 tp->rack.reo_wnd_steps = 1; 2674 tp->rack.last_delivered = 0; 2675 tp->rack.reo_wnd_persist = 0; 2676 tp->rack.dsack_seen = 0; 2677 tp->syn_data_acked = 0; 2678 tp->rx_opt.saw_tstamp = 0; 2679 tp->rx_opt.dsack = 0; 2680 tp->rx_opt.num_sacks = 0; 2681 tp->rcv_ooopack = 0; 2682 2683 2684 /* Clean up fastopen related fields */ 2685 tcp_free_fastopen_req(tp); 2686 inet->defer_connect = 0; 2687 tp->fastopen_client_fail = 0; 2688 2689 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 2690 2691 if (sk->sk_frag.page) { 2692 put_page(sk->sk_frag.page); 2693 sk->sk_frag.page = NULL; 2694 sk->sk_frag.offset = 0; 2695 } 2696 2697 sk->sk_error_report(sk); 2698 return 0; 2699 } 2700 EXPORT_SYMBOL(tcp_disconnect); 2701 2702 static inline bool tcp_can_repair_sock(const struct sock *sk) 2703 { 2704 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) && 2705 (sk->sk_state != TCP_LISTEN); 2706 } 2707 2708 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len) 2709 { 2710 struct tcp_repair_window opt; 2711 2712 if (!tp->repair) 2713 return -EPERM; 2714 2715 if (len != sizeof(opt)) 2716 return -EINVAL; 2717 2718 if (copy_from_user(&opt, optbuf, sizeof(opt))) 2719 return -EFAULT; 2720 2721 if (opt.max_window < opt.snd_wnd) 2722 return -EINVAL; 2723 2724 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd)) 2725 return -EINVAL; 2726 2727 if (after(opt.rcv_wup, tp->rcv_nxt)) 2728 return -EINVAL; 2729 2730 tp->snd_wl1 = opt.snd_wl1; 2731 tp->snd_wnd = opt.snd_wnd; 2732 tp->max_window = opt.max_window; 2733 2734 tp->rcv_wnd = opt.rcv_wnd; 2735 tp->rcv_wup = opt.rcv_wup; 2736 2737 return 0; 2738 } 2739 2740 static int tcp_repair_options_est(struct sock *sk, 2741 struct tcp_repair_opt __user *optbuf, unsigned int len) 2742 { 2743 struct tcp_sock *tp = tcp_sk(sk); 2744 struct tcp_repair_opt opt; 2745 2746 while (len >= sizeof(opt)) { 2747 if (copy_from_user(&opt, optbuf, sizeof(opt))) 2748 return -EFAULT; 2749 2750 optbuf++; 2751 len -= sizeof(opt); 2752 2753 switch (opt.opt_code) { 2754 case TCPOPT_MSS: 2755 tp->rx_opt.mss_clamp = opt.opt_val; 2756 tcp_mtup_init(sk); 2757 break; 2758 case TCPOPT_WINDOW: 2759 { 2760 u16 snd_wscale = opt.opt_val & 0xFFFF; 2761 u16 rcv_wscale = opt.opt_val >> 16; 2762 2763 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE) 2764 return -EFBIG; 2765 2766 tp->rx_opt.snd_wscale = snd_wscale; 2767 tp->rx_opt.rcv_wscale = rcv_wscale; 2768 tp->rx_opt.wscale_ok = 1; 2769 } 2770 break; 2771 case TCPOPT_SACK_PERM: 2772 if (opt.opt_val != 0) 2773 return -EINVAL; 2774 2775 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 2776 break; 2777 case TCPOPT_TIMESTAMP: 2778 if (opt.opt_val != 0) 2779 return -EINVAL; 2780 2781 tp->rx_opt.tstamp_ok = 1; 2782 break; 2783 } 2784 } 2785 2786 return 0; 2787 } 2788 2789 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled); 2790 EXPORT_SYMBOL(tcp_tx_delay_enabled); 2791 2792 static void tcp_enable_tx_delay(void) 2793 { 2794 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) { 2795 static int __tcp_tx_delay_enabled = 0; 2796 2797 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) { 2798 static_branch_enable(&tcp_tx_delay_enabled); 2799 pr_info("TCP_TX_DELAY enabled\n"); 2800 } 2801 } 2802 } 2803 2804 /* 2805 * Socket option code for TCP. 2806 */ 2807 static int do_tcp_setsockopt(struct sock *sk, int level, 2808 int optname, char __user *optval, unsigned int optlen) 2809 { 2810 struct tcp_sock *tp = tcp_sk(sk); 2811 struct inet_connection_sock *icsk = inet_csk(sk); 2812 struct net *net = sock_net(sk); 2813 int val; 2814 int err = 0; 2815 2816 /* These are data/string values, all the others are ints */ 2817 switch (optname) { 2818 case TCP_CONGESTION: { 2819 char name[TCP_CA_NAME_MAX]; 2820 2821 if (optlen < 1) 2822 return -EINVAL; 2823 2824 val = strncpy_from_user(name, optval, 2825 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 2826 if (val < 0) 2827 return -EFAULT; 2828 name[val] = 0; 2829 2830 lock_sock(sk); 2831 err = tcp_set_congestion_control(sk, name, true, true, 2832 ns_capable(sock_net(sk)->user_ns, 2833 CAP_NET_ADMIN)); 2834 release_sock(sk); 2835 return err; 2836 } 2837 case TCP_ULP: { 2838 char name[TCP_ULP_NAME_MAX]; 2839 2840 if (optlen < 1) 2841 return -EINVAL; 2842 2843 val = strncpy_from_user(name, optval, 2844 min_t(long, TCP_ULP_NAME_MAX - 1, 2845 optlen)); 2846 if (val < 0) 2847 return -EFAULT; 2848 name[val] = 0; 2849 2850 lock_sock(sk); 2851 err = tcp_set_ulp(sk, name); 2852 release_sock(sk); 2853 return err; 2854 } 2855 case TCP_FASTOPEN_KEY: { 2856 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH]; 2857 __u8 *backup_key = NULL; 2858 2859 /* Allow a backup key as well to facilitate key rotation 2860 * First key is the active one. 2861 */ 2862 if (optlen != TCP_FASTOPEN_KEY_LENGTH && 2863 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH) 2864 return -EINVAL; 2865 2866 if (copy_from_user(key, optval, optlen)) 2867 return -EFAULT; 2868 2869 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH) 2870 backup_key = key + TCP_FASTOPEN_KEY_LENGTH; 2871 2872 return tcp_fastopen_reset_cipher(net, sk, key, backup_key); 2873 } 2874 default: 2875 /* fallthru */ 2876 break; 2877 } 2878 2879 if (optlen < sizeof(int)) 2880 return -EINVAL; 2881 2882 if (get_user(val, (int __user *)optval)) 2883 return -EFAULT; 2884 2885 lock_sock(sk); 2886 2887 switch (optname) { 2888 case TCP_MAXSEG: 2889 /* Values greater than interface MTU won't take effect. However 2890 * at the point when this call is done we typically don't yet 2891 * know which interface is going to be used 2892 */ 2893 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) { 2894 err = -EINVAL; 2895 break; 2896 } 2897 tp->rx_opt.user_mss = val; 2898 break; 2899 2900 case TCP_NODELAY: 2901 if (val) { 2902 /* TCP_NODELAY is weaker than TCP_CORK, so that 2903 * this option on corked socket is remembered, but 2904 * it is not activated until cork is cleared. 2905 * 2906 * However, when TCP_NODELAY is set we make 2907 * an explicit push, which overrides even TCP_CORK 2908 * for currently queued segments. 2909 */ 2910 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 2911 tcp_push_pending_frames(sk); 2912 } else { 2913 tp->nonagle &= ~TCP_NAGLE_OFF; 2914 } 2915 break; 2916 2917 case TCP_THIN_LINEAR_TIMEOUTS: 2918 if (val < 0 || val > 1) 2919 err = -EINVAL; 2920 else 2921 tp->thin_lto = val; 2922 break; 2923 2924 case TCP_THIN_DUPACK: 2925 if (val < 0 || val > 1) 2926 err = -EINVAL; 2927 break; 2928 2929 case TCP_REPAIR: 2930 if (!tcp_can_repair_sock(sk)) 2931 err = -EPERM; 2932 else if (val == TCP_REPAIR_ON) { 2933 tp->repair = 1; 2934 sk->sk_reuse = SK_FORCE_REUSE; 2935 tp->repair_queue = TCP_NO_QUEUE; 2936 } else if (val == TCP_REPAIR_OFF) { 2937 tp->repair = 0; 2938 sk->sk_reuse = SK_NO_REUSE; 2939 tcp_send_window_probe(sk); 2940 } else if (val == TCP_REPAIR_OFF_NO_WP) { 2941 tp->repair = 0; 2942 sk->sk_reuse = SK_NO_REUSE; 2943 } else 2944 err = -EINVAL; 2945 2946 break; 2947 2948 case TCP_REPAIR_QUEUE: 2949 if (!tp->repair) 2950 err = -EPERM; 2951 else if ((unsigned int)val < TCP_QUEUES_NR) 2952 tp->repair_queue = val; 2953 else 2954 err = -EINVAL; 2955 break; 2956 2957 case TCP_QUEUE_SEQ: 2958 if (sk->sk_state != TCP_CLOSE) 2959 err = -EPERM; 2960 else if (tp->repair_queue == TCP_SEND_QUEUE) 2961 WRITE_ONCE(tp->write_seq, val); 2962 else if (tp->repair_queue == TCP_RECV_QUEUE) { 2963 WRITE_ONCE(tp->rcv_nxt, val); 2964 WRITE_ONCE(tp->copied_seq, val); 2965 } 2966 else 2967 err = -EINVAL; 2968 break; 2969 2970 case TCP_REPAIR_OPTIONS: 2971 if (!tp->repair) 2972 err = -EINVAL; 2973 else if (sk->sk_state == TCP_ESTABLISHED) 2974 err = tcp_repair_options_est(sk, 2975 (struct tcp_repair_opt __user *)optval, 2976 optlen); 2977 else 2978 err = -EPERM; 2979 break; 2980 2981 case TCP_CORK: 2982 /* When set indicates to always queue non-full frames. 2983 * Later the user clears this option and we transmit 2984 * any pending partial frames in the queue. This is 2985 * meant to be used alongside sendfile() to get properly 2986 * filled frames when the user (for example) must write 2987 * out headers with a write() call first and then use 2988 * sendfile to send out the data parts. 2989 * 2990 * TCP_CORK can be set together with TCP_NODELAY and it is 2991 * stronger than TCP_NODELAY. 2992 */ 2993 if (val) { 2994 tp->nonagle |= TCP_NAGLE_CORK; 2995 } else { 2996 tp->nonagle &= ~TCP_NAGLE_CORK; 2997 if (tp->nonagle&TCP_NAGLE_OFF) 2998 tp->nonagle |= TCP_NAGLE_PUSH; 2999 tcp_push_pending_frames(sk); 3000 } 3001 break; 3002 3003 case TCP_KEEPIDLE: 3004 if (val < 1 || val > MAX_TCP_KEEPIDLE) 3005 err = -EINVAL; 3006 else { 3007 tp->keepalive_time = val * HZ; 3008 if (sock_flag(sk, SOCK_KEEPOPEN) && 3009 !((1 << sk->sk_state) & 3010 (TCPF_CLOSE | TCPF_LISTEN))) { 3011 u32 elapsed = keepalive_time_elapsed(tp); 3012 if (tp->keepalive_time > elapsed) 3013 elapsed = tp->keepalive_time - elapsed; 3014 else 3015 elapsed = 0; 3016 inet_csk_reset_keepalive_timer(sk, elapsed); 3017 } 3018 } 3019 break; 3020 case TCP_KEEPINTVL: 3021 if (val < 1 || val > MAX_TCP_KEEPINTVL) 3022 err = -EINVAL; 3023 else 3024 tp->keepalive_intvl = val * HZ; 3025 break; 3026 case TCP_KEEPCNT: 3027 if (val < 1 || val > MAX_TCP_KEEPCNT) 3028 err = -EINVAL; 3029 else 3030 tp->keepalive_probes = val; 3031 break; 3032 case TCP_SYNCNT: 3033 if (val < 1 || val > MAX_TCP_SYNCNT) 3034 err = -EINVAL; 3035 else 3036 icsk->icsk_syn_retries = val; 3037 break; 3038 3039 case TCP_SAVE_SYN: 3040 if (val < 0 || val > 1) 3041 err = -EINVAL; 3042 else 3043 tp->save_syn = val; 3044 break; 3045 3046 case TCP_LINGER2: 3047 if (val < 0) 3048 tp->linger2 = -1; 3049 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ) 3050 tp->linger2 = 0; 3051 else 3052 tp->linger2 = val * HZ; 3053 break; 3054 3055 case TCP_DEFER_ACCEPT: 3056 /* Translate value in seconds to number of retransmits */ 3057 icsk->icsk_accept_queue.rskq_defer_accept = 3058 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 3059 TCP_RTO_MAX / HZ); 3060 break; 3061 3062 case TCP_WINDOW_CLAMP: 3063 if (!val) { 3064 if (sk->sk_state != TCP_CLOSE) { 3065 err = -EINVAL; 3066 break; 3067 } 3068 tp->window_clamp = 0; 3069 } else 3070 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? 3071 SOCK_MIN_RCVBUF / 2 : val; 3072 break; 3073 3074 case TCP_QUICKACK: 3075 if (!val) { 3076 inet_csk_enter_pingpong_mode(sk); 3077 } else { 3078 inet_csk_exit_pingpong_mode(sk); 3079 if ((1 << sk->sk_state) & 3080 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 3081 inet_csk_ack_scheduled(sk)) { 3082 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 3083 tcp_cleanup_rbuf(sk, 1); 3084 if (!(val & 1)) 3085 inet_csk_enter_pingpong_mode(sk); 3086 } 3087 } 3088 break; 3089 3090 #ifdef CONFIG_TCP_MD5SIG 3091 case TCP_MD5SIG: 3092 case TCP_MD5SIG_EXT: 3093 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 3094 err = tp->af_specific->md5_parse(sk, optname, optval, optlen); 3095 else 3096 err = -EINVAL; 3097 break; 3098 #endif 3099 case TCP_USER_TIMEOUT: 3100 /* Cap the max time in ms TCP will retry or probe the window 3101 * before giving up and aborting (ETIMEDOUT) a connection. 3102 */ 3103 if (val < 0) 3104 err = -EINVAL; 3105 else 3106 icsk->icsk_user_timeout = val; 3107 break; 3108 3109 case TCP_FASTOPEN: 3110 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE | 3111 TCPF_LISTEN))) { 3112 tcp_fastopen_init_key_once(net); 3113 3114 fastopen_queue_tune(sk, val); 3115 } else { 3116 err = -EINVAL; 3117 } 3118 break; 3119 case TCP_FASTOPEN_CONNECT: 3120 if (val > 1 || val < 0) { 3121 err = -EINVAL; 3122 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) { 3123 if (sk->sk_state == TCP_CLOSE) 3124 tp->fastopen_connect = val; 3125 else 3126 err = -EINVAL; 3127 } else { 3128 err = -EOPNOTSUPP; 3129 } 3130 break; 3131 case TCP_FASTOPEN_NO_COOKIE: 3132 if (val > 1 || val < 0) 3133 err = -EINVAL; 3134 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 3135 err = -EINVAL; 3136 else 3137 tp->fastopen_no_cookie = val; 3138 break; 3139 case TCP_TIMESTAMP: 3140 if (!tp->repair) 3141 err = -EPERM; 3142 else 3143 tp->tsoffset = val - tcp_time_stamp_raw(); 3144 break; 3145 case TCP_REPAIR_WINDOW: 3146 err = tcp_repair_set_window(tp, optval, optlen); 3147 break; 3148 case TCP_NOTSENT_LOWAT: 3149 tp->notsent_lowat = val; 3150 sk->sk_write_space(sk); 3151 break; 3152 case TCP_INQ: 3153 if (val > 1 || val < 0) 3154 err = -EINVAL; 3155 else 3156 tp->recvmsg_inq = val; 3157 break; 3158 case TCP_TX_DELAY: 3159 if (val) 3160 tcp_enable_tx_delay(); 3161 tp->tcp_tx_delay = val; 3162 break; 3163 default: 3164 err = -ENOPROTOOPT; 3165 break; 3166 } 3167 3168 release_sock(sk); 3169 return err; 3170 } 3171 3172 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval, 3173 unsigned int optlen) 3174 { 3175 const struct inet_connection_sock *icsk = inet_csk(sk); 3176 3177 if (level != SOL_TCP) 3178 return icsk->icsk_af_ops->setsockopt(sk, level, optname, 3179 optval, optlen); 3180 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 3181 } 3182 EXPORT_SYMBOL(tcp_setsockopt); 3183 3184 #ifdef CONFIG_COMPAT 3185 int compat_tcp_setsockopt(struct sock *sk, int level, int optname, 3186 char __user *optval, unsigned int optlen) 3187 { 3188 if (level != SOL_TCP) 3189 return inet_csk_compat_setsockopt(sk, level, optname, 3190 optval, optlen); 3191 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 3192 } 3193 EXPORT_SYMBOL(compat_tcp_setsockopt); 3194 #endif 3195 3196 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp, 3197 struct tcp_info *info) 3198 { 3199 u64 stats[__TCP_CHRONO_MAX], total = 0; 3200 enum tcp_chrono i; 3201 3202 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) { 3203 stats[i] = tp->chrono_stat[i - 1]; 3204 if (i == tp->chrono_type) 3205 stats[i] += tcp_jiffies32 - tp->chrono_start; 3206 stats[i] *= USEC_PER_SEC / HZ; 3207 total += stats[i]; 3208 } 3209 3210 info->tcpi_busy_time = total; 3211 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED]; 3212 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED]; 3213 } 3214 3215 /* Return information about state of tcp endpoint in API format. */ 3216 void tcp_get_info(struct sock *sk, struct tcp_info *info) 3217 { 3218 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */ 3219 const struct inet_connection_sock *icsk = inet_csk(sk); 3220 unsigned long rate; 3221 u32 now; 3222 u64 rate64; 3223 bool slow; 3224 3225 memset(info, 0, sizeof(*info)); 3226 if (sk->sk_type != SOCK_STREAM) 3227 return; 3228 3229 info->tcpi_state = inet_sk_state_load(sk); 3230 3231 /* Report meaningful fields for all TCP states, including listeners */ 3232 rate = READ_ONCE(sk->sk_pacing_rate); 3233 rate64 = (rate != ~0UL) ? rate : ~0ULL; 3234 info->tcpi_pacing_rate = rate64; 3235 3236 rate = READ_ONCE(sk->sk_max_pacing_rate); 3237 rate64 = (rate != ~0UL) ? rate : ~0ULL; 3238 info->tcpi_max_pacing_rate = rate64; 3239 3240 info->tcpi_reordering = tp->reordering; 3241 info->tcpi_snd_cwnd = tp->snd_cwnd; 3242 3243 if (info->tcpi_state == TCP_LISTEN) { 3244 /* listeners aliased fields : 3245 * tcpi_unacked -> Number of children ready for accept() 3246 * tcpi_sacked -> max backlog 3247 */ 3248 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog); 3249 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog); 3250 return; 3251 } 3252 3253 slow = lock_sock_fast(sk); 3254 3255 info->tcpi_ca_state = icsk->icsk_ca_state; 3256 info->tcpi_retransmits = icsk->icsk_retransmits; 3257 info->tcpi_probes = icsk->icsk_probes_out; 3258 info->tcpi_backoff = icsk->icsk_backoff; 3259 3260 if (tp->rx_opt.tstamp_ok) 3261 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 3262 if (tcp_is_sack(tp)) 3263 info->tcpi_options |= TCPI_OPT_SACK; 3264 if (tp->rx_opt.wscale_ok) { 3265 info->tcpi_options |= TCPI_OPT_WSCALE; 3266 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 3267 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 3268 } 3269 3270 if (tp->ecn_flags & TCP_ECN_OK) 3271 info->tcpi_options |= TCPI_OPT_ECN; 3272 if (tp->ecn_flags & TCP_ECN_SEEN) 3273 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 3274 if (tp->syn_data_acked) 3275 info->tcpi_options |= TCPI_OPT_SYN_DATA; 3276 3277 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 3278 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato); 3279 info->tcpi_snd_mss = tp->mss_cache; 3280 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 3281 3282 info->tcpi_unacked = tp->packets_out; 3283 info->tcpi_sacked = tp->sacked_out; 3284 3285 info->tcpi_lost = tp->lost_out; 3286 info->tcpi_retrans = tp->retrans_out; 3287 3288 now = tcp_jiffies32; 3289 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 3290 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 3291 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 3292 3293 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 3294 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 3295 info->tcpi_rtt = tp->srtt_us >> 3; 3296 info->tcpi_rttvar = tp->mdev_us >> 2; 3297 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 3298 info->tcpi_advmss = tp->advmss; 3299 3300 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3; 3301 info->tcpi_rcv_space = tp->rcvq_space.space; 3302 3303 info->tcpi_total_retrans = tp->total_retrans; 3304 3305 info->tcpi_bytes_acked = tp->bytes_acked; 3306 info->tcpi_bytes_received = tp->bytes_received; 3307 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt); 3308 tcp_get_info_chrono_stats(tp, info); 3309 3310 info->tcpi_segs_out = tp->segs_out; 3311 info->tcpi_segs_in = tp->segs_in; 3312 3313 info->tcpi_min_rtt = tcp_min_rtt(tp); 3314 info->tcpi_data_segs_in = tp->data_segs_in; 3315 info->tcpi_data_segs_out = tp->data_segs_out; 3316 3317 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0; 3318 rate64 = tcp_compute_delivery_rate(tp); 3319 if (rate64) 3320 info->tcpi_delivery_rate = rate64; 3321 info->tcpi_delivered = tp->delivered; 3322 info->tcpi_delivered_ce = tp->delivered_ce; 3323 info->tcpi_bytes_sent = tp->bytes_sent; 3324 info->tcpi_bytes_retrans = tp->bytes_retrans; 3325 info->tcpi_dsack_dups = tp->dsack_dups; 3326 info->tcpi_reord_seen = tp->reord_seen; 3327 info->tcpi_rcv_ooopack = tp->rcv_ooopack; 3328 info->tcpi_snd_wnd = tp->snd_wnd; 3329 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail; 3330 unlock_sock_fast(sk, slow); 3331 } 3332 EXPORT_SYMBOL_GPL(tcp_get_info); 3333 3334 static size_t tcp_opt_stats_get_size(void) 3335 { 3336 return 3337 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */ 3338 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */ 3339 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */ 3340 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */ 3341 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */ 3342 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */ 3343 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */ 3344 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */ 3345 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */ 3346 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */ 3347 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */ 3348 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */ 3349 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */ 3350 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */ 3351 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */ 3352 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */ 3353 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */ 3354 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */ 3355 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */ 3356 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */ 3357 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */ 3358 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */ 3359 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */ 3360 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */ 3361 0; 3362 } 3363 3364 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk) 3365 { 3366 const struct tcp_sock *tp = tcp_sk(sk); 3367 struct sk_buff *stats; 3368 struct tcp_info info; 3369 unsigned long rate; 3370 u64 rate64; 3371 3372 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC); 3373 if (!stats) 3374 return NULL; 3375 3376 tcp_get_info_chrono_stats(tp, &info); 3377 nla_put_u64_64bit(stats, TCP_NLA_BUSY, 3378 info.tcpi_busy_time, TCP_NLA_PAD); 3379 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED, 3380 info.tcpi_rwnd_limited, TCP_NLA_PAD); 3381 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED, 3382 info.tcpi_sndbuf_limited, TCP_NLA_PAD); 3383 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT, 3384 tp->data_segs_out, TCP_NLA_PAD); 3385 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS, 3386 tp->total_retrans, TCP_NLA_PAD); 3387 3388 rate = READ_ONCE(sk->sk_pacing_rate); 3389 rate64 = (rate != ~0UL) ? rate : ~0ULL; 3390 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD); 3391 3392 rate64 = tcp_compute_delivery_rate(tp); 3393 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD); 3394 3395 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd); 3396 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering); 3397 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp)); 3398 3399 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits); 3400 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited); 3401 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh); 3402 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered); 3403 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce); 3404 3405 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una); 3406 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state); 3407 3408 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent, 3409 TCP_NLA_PAD); 3410 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans, 3411 TCP_NLA_PAD); 3412 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups); 3413 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen); 3414 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3); 3415 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash); 3416 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT, 3417 max_t(int, 0, tp->write_seq - tp->snd_nxt)); 3418 3419 return stats; 3420 } 3421 3422 static int do_tcp_getsockopt(struct sock *sk, int level, 3423 int optname, char __user *optval, int __user *optlen) 3424 { 3425 struct inet_connection_sock *icsk = inet_csk(sk); 3426 struct tcp_sock *tp = tcp_sk(sk); 3427 struct net *net = sock_net(sk); 3428 int val, len; 3429 3430 if (get_user(len, optlen)) 3431 return -EFAULT; 3432 3433 len = min_t(unsigned int, len, sizeof(int)); 3434 3435 if (len < 0) 3436 return -EINVAL; 3437 3438 switch (optname) { 3439 case TCP_MAXSEG: 3440 val = tp->mss_cache; 3441 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 3442 val = tp->rx_opt.user_mss; 3443 if (tp->repair) 3444 val = tp->rx_opt.mss_clamp; 3445 break; 3446 case TCP_NODELAY: 3447 val = !!(tp->nonagle&TCP_NAGLE_OFF); 3448 break; 3449 case TCP_CORK: 3450 val = !!(tp->nonagle&TCP_NAGLE_CORK); 3451 break; 3452 case TCP_KEEPIDLE: 3453 val = keepalive_time_when(tp) / HZ; 3454 break; 3455 case TCP_KEEPINTVL: 3456 val = keepalive_intvl_when(tp) / HZ; 3457 break; 3458 case TCP_KEEPCNT: 3459 val = keepalive_probes(tp); 3460 break; 3461 case TCP_SYNCNT: 3462 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries; 3463 break; 3464 case TCP_LINGER2: 3465 val = tp->linger2; 3466 if (val >= 0) 3467 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ; 3468 break; 3469 case TCP_DEFER_ACCEPT: 3470 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept, 3471 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ); 3472 break; 3473 case TCP_WINDOW_CLAMP: 3474 val = tp->window_clamp; 3475 break; 3476 case TCP_INFO: { 3477 struct tcp_info info; 3478 3479 if (get_user(len, optlen)) 3480 return -EFAULT; 3481 3482 tcp_get_info(sk, &info); 3483 3484 len = min_t(unsigned int, len, sizeof(info)); 3485 if (put_user(len, optlen)) 3486 return -EFAULT; 3487 if (copy_to_user(optval, &info, len)) 3488 return -EFAULT; 3489 return 0; 3490 } 3491 case TCP_CC_INFO: { 3492 const struct tcp_congestion_ops *ca_ops; 3493 union tcp_cc_info info; 3494 size_t sz = 0; 3495 int attr; 3496 3497 if (get_user(len, optlen)) 3498 return -EFAULT; 3499 3500 ca_ops = icsk->icsk_ca_ops; 3501 if (ca_ops && ca_ops->get_info) 3502 sz = ca_ops->get_info(sk, ~0U, &attr, &info); 3503 3504 len = min_t(unsigned int, len, sz); 3505 if (put_user(len, optlen)) 3506 return -EFAULT; 3507 if (copy_to_user(optval, &info, len)) 3508 return -EFAULT; 3509 return 0; 3510 } 3511 case TCP_QUICKACK: 3512 val = !inet_csk_in_pingpong_mode(sk); 3513 break; 3514 3515 case TCP_CONGESTION: 3516 if (get_user(len, optlen)) 3517 return -EFAULT; 3518 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 3519 if (put_user(len, optlen)) 3520 return -EFAULT; 3521 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len)) 3522 return -EFAULT; 3523 return 0; 3524 3525 case TCP_ULP: 3526 if (get_user(len, optlen)) 3527 return -EFAULT; 3528 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX); 3529 if (!icsk->icsk_ulp_ops) { 3530 if (put_user(0, optlen)) 3531 return -EFAULT; 3532 return 0; 3533 } 3534 if (put_user(len, optlen)) 3535 return -EFAULT; 3536 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len)) 3537 return -EFAULT; 3538 return 0; 3539 3540 case TCP_FASTOPEN_KEY: { 3541 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH]; 3542 struct tcp_fastopen_context *ctx; 3543 unsigned int key_len = 0; 3544 3545 if (get_user(len, optlen)) 3546 return -EFAULT; 3547 3548 rcu_read_lock(); 3549 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx); 3550 if (ctx) { 3551 key_len = tcp_fastopen_context_len(ctx) * 3552 TCP_FASTOPEN_KEY_LENGTH; 3553 memcpy(&key[0], &ctx->key[0], key_len); 3554 } 3555 rcu_read_unlock(); 3556 3557 len = min_t(unsigned int, len, key_len); 3558 if (put_user(len, optlen)) 3559 return -EFAULT; 3560 if (copy_to_user(optval, key, len)) 3561 return -EFAULT; 3562 return 0; 3563 } 3564 case TCP_THIN_LINEAR_TIMEOUTS: 3565 val = tp->thin_lto; 3566 break; 3567 3568 case TCP_THIN_DUPACK: 3569 val = 0; 3570 break; 3571 3572 case TCP_REPAIR: 3573 val = tp->repair; 3574 break; 3575 3576 case TCP_REPAIR_QUEUE: 3577 if (tp->repair) 3578 val = tp->repair_queue; 3579 else 3580 return -EINVAL; 3581 break; 3582 3583 case TCP_REPAIR_WINDOW: { 3584 struct tcp_repair_window opt; 3585 3586 if (get_user(len, optlen)) 3587 return -EFAULT; 3588 3589 if (len != sizeof(opt)) 3590 return -EINVAL; 3591 3592 if (!tp->repair) 3593 return -EPERM; 3594 3595 opt.snd_wl1 = tp->snd_wl1; 3596 opt.snd_wnd = tp->snd_wnd; 3597 opt.max_window = tp->max_window; 3598 opt.rcv_wnd = tp->rcv_wnd; 3599 opt.rcv_wup = tp->rcv_wup; 3600 3601 if (copy_to_user(optval, &opt, len)) 3602 return -EFAULT; 3603 return 0; 3604 } 3605 case TCP_QUEUE_SEQ: 3606 if (tp->repair_queue == TCP_SEND_QUEUE) 3607 val = tp->write_seq; 3608 else if (tp->repair_queue == TCP_RECV_QUEUE) 3609 val = tp->rcv_nxt; 3610 else 3611 return -EINVAL; 3612 break; 3613 3614 case TCP_USER_TIMEOUT: 3615 val = icsk->icsk_user_timeout; 3616 break; 3617 3618 case TCP_FASTOPEN: 3619 val = icsk->icsk_accept_queue.fastopenq.max_qlen; 3620 break; 3621 3622 case TCP_FASTOPEN_CONNECT: 3623 val = tp->fastopen_connect; 3624 break; 3625 3626 case TCP_FASTOPEN_NO_COOKIE: 3627 val = tp->fastopen_no_cookie; 3628 break; 3629 3630 case TCP_TX_DELAY: 3631 val = tp->tcp_tx_delay; 3632 break; 3633 3634 case TCP_TIMESTAMP: 3635 val = tcp_time_stamp_raw() + tp->tsoffset; 3636 break; 3637 case TCP_NOTSENT_LOWAT: 3638 val = tp->notsent_lowat; 3639 break; 3640 case TCP_INQ: 3641 val = tp->recvmsg_inq; 3642 break; 3643 case TCP_SAVE_SYN: 3644 val = tp->save_syn; 3645 break; 3646 case TCP_SAVED_SYN: { 3647 if (get_user(len, optlen)) 3648 return -EFAULT; 3649 3650 lock_sock(sk); 3651 if (tp->saved_syn) { 3652 if (len < tp->saved_syn[0]) { 3653 if (put_user(tp->saved_syn[0], optlen)) { 3654 release_sock(sk); 3655 return -EFAULT; 3656 } 3657 release_sock(sk); 3658 return -EINVAL; 3659 } 3660 len = tp->saved_syn[0]; 3661 if (put_user(len, optlen)) { 3662 release_sock(sk); 3663 return -EFAULT; 3664 } 3665 if (copy_to_user(optval, tp->saved_syn + 1, len)) { 3666 release_sock(sk); 3667 return -EFAULT; 3668 } 3669 tcp_saved_syn_free(tp); 3670 release_sock(sk); 3671 } else { 3672 release_sock(sk); 3673 len = 0; 3674 if (put_user(len, optlen)) 3675 return -EFAULT; 3676 } 3677 return 0; 3678 } 3679 #ifdef CONFIG_MMU 3680 case TCP_ZEROCOPY_RECEIVE: { 3681 struct tcp_zerocopy_receive zc; 3682 int err; 3683 3684 if (get_user(len, optlen)) 3685 return -EFAULT; 3686 if (len < offsetofend(struct tcp_zerocopy_receive, length)) 3687 return -EINVAL; 3688 if (len > sizeof(zc)) { 3689 len = sizeof(zc); 3690 if (put_user(len, optlen)) 3691 return -EFAULT; 3692 } 3693 if (copy_from_user(&zc, optval, len)) 3694 return -EFAULT; 3695 lock_sock(sk); 3696 err = tcp_zerocopy_receive(sk, &zc); 3697 release_sock(sk); 3698 if (len == sizeof(zc)) 3699 goto zerocopy_rcv_sk_err; 3700 switch (len) { 3701 case offsetofend(struct tcp_zerocopy_receive, err): 3702 goto zerocopy_rcv_sk_err; 3703 case offsetofend(struct tcp_zerocopy_receive, inq): 3704 goto zerocopy_rcv_inq; 3705 case offsetofend(struct tcp_zerocopy_receive, length): 3706 default: 3707 goto zerocopy_rcv_out; 3708 } 3709 zerocopy_rcv_sk_err: 3710 if (!err) 3711 zc.err = sock_error(sk); 3712 zerocopy_rcv_inq: 3713 zc.inq = tcp_inq_hint(sk); 3714 zerocopy_rcv_out: 3715 if (!err && copy_to_user(optval, &zc, len)) 3716 err = -EFAULT; 3717 return err; 3718 } 3719 #endif 3720 default: 3721 return -ENOPROTOOPT; 3722 } 3723 3724 if (put_user(len, optlen)) 3725 return -EFAULT; 3726 if (copy_to_user(optval, &val, len)) 3727 return -EFAULT; 3728 return 0; 3729 } 3730 3731 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 3732 int __user *optlen) 3733 { 3734 struct inet_connection_sock *icsk = inet_csk(sk); 3735 3736 if (level != SOL_TCP) 3737 return icsk->icsk_af_ops->getsockopt(sk, level, optname, 3738 optval, optlen); 3739 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 3740 } 3741 EXPORT_SYMBOL(tcp_getsockopt); 3742 3743 #ifdef CONFIG_COMPAT 3744 int compat_tcp_getsockopt(struct sock *sk, int level, int optname, 3745 char __user *optval, int __user *optlen) 3746 { 3747 if (level != SOL_TCP) 3748 return inet_csk_compat_getsockopt(sk, level, optname, 3749 optval, optlen); 3750 return do_tcp_getsockopt(sk, level, optname, optval, optlen); 3751 } 3752 EXPORT_SYMBOL(compat_tcp_getsockopt); 3753 #endif 3754 3755 #ifdef CONFIG_TCP_MD5SIG 3756 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool); 3757 static DEFINE_MUTEX(tcp_md5sig_mutex); 3758 static bool tcp_md5sig_pool_populated = false; 3759 3760 static void __tcp_alloc_md5sig_pool(void) 3761 { 3762 struct crypto_ahash *hash; 3763 int cpu; 3764 3765 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC); 3766 if (IS_ERR(hash)) 3767 return; 3768 3769 for_each_possible_cpu(cpu) { 3770 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch; 3771 struct ahash_request *req; 3772 3773 if (!scratch) { 3774 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) + 3775 sizeof(struct tcphdr), 3776 GFP_KERNEL, 3777 cpu_to_node(cpu)); 3778 if (!scratch) 3779 return; 3780 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch; 3781 } 3782 if (per_cpu(tcp_md5sig_pool, cpu).md5_req) 3783 continue; 3784 3785 req = ahash_request_alloc(hash, GFP_KERNEL); 3786 if (!req) 3787 return; 3788 3789 ahash_request_set_callback(req, 0, NULL, NULL); 3790 3791 per_cpu(tcp_md5sig_pool, cpu).md5_req = req; 3792 } 3793 /* before setting tcp_md5sig_pool_populated, we must commit all writes 3794 * to memory. See smp_rmb() in tcp_get_md5sig_pool() 3795 */ 3796 smp_wmb(); 3797 tcp_md5sig_pool_populated = true; 3798 } 3799 3800 bool tcp_alloc_md5sig_pool(void) 3801 { 3802 if (unlikely(!tcp_md5sig_pool_populated)) { 3803 mutex_lock(&tcp_md5sig_mutex); 3804 3805 if (!tcp_md5sig_pool_populated) { 3806 __tcp_alloc_md5sig_pool(); 3807 if (tcp_md5sig_pool_populated) 3808 static_branch_inc(&tcp_md5_needed); 3809 } 3810 3811 mutex_unlock(&tcp_md5sig_mutex); 3812 } 3813 return tcp_md5sig_pool_populated; 3814 } 3815 EXPORT_SYMBOL(tcp_alloc_md5sig_pool); 3816 3817 3818 /** 3819 * tcp_get_md5sig_pool - get md5sig_pool for this user 3820 * 3821 * We use percpu structure, so if we succeed, we exit with preemption 3822 * and BH disabled, to make sure another thread or softirq handling 3823 * wont try to get same context. 3824 */ 3825 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void) 3826 { 3827 local_bh_disable(); 3828 3829 if (tcp_md5sig_pool_populated) { 3830 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */ 3831 smp_rmb(); 3832 return this_cpu_ptr(&tcp_md5sig_pool); 3833 } 3834 local_bh_enable(); 3835 return NULL; 3836 } 3837 EXPORT_SYMBOL(tcp_get_md5sig_pool); 3838 3839 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp, 3840 const struct sk_buff *skb, unsigned int header_len) 3841 { 3842 struct scatterlist sg; 3843 const struct tcphdr *tp = tcp_hdr(skb); 3844 struct ahash_request *req = hp->md5_req; 3845 unsigned int i; 3846 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 3847 skb_headlen(skb) - header_len : 0; 3848 const struct skb_shared_info *shi = skb_shinfo(skb); 3849 struct sk_buff *frag_iter; 3850 3851 sg_init_table(&sg, 1); 3852 3853 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len); 3854 ahash_request_set_crypt(req, &sg, NULL, head_data_len); 3855 if (crypto_ahash_update(req)) 3856 return 1; 3857 3858 for (i = 0; i < shi->nr_frags; ++i) { 3859 const skb_frag_t *f = &shi->frags[i]; 3860 unsigned int offset = skb_frag_off(f); 3861 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT); 3862 3863 sg_set_page(&sg, page, skb_frag_size(f), 3864 offset_in_page(offset)); 3865 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f)); 3866 if (crypto_ahash_update(req)) 3867 return 1; 3868 } 3869 3870 skb_walk_frags(skb, frag_iter) 3871 if (tcp_md5_hash_skb_data(hp, frag_iter, 0)) 3872 return 1; 3873 3874 return 0; 3875 } 3876 EXPORT_SYMBOL(tcp_md5_hash_skb_data); 3877 3878 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key) 3879 { 3880 struct scatterlist sg; 3881 3882 sg_init_one(&sg, key->key, key->keylen); 3883 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen); 3884 return crypto_ahash_update(hp->md5_req); 3885 } 3886 EXPORT_SYMBOL(tcp_md5_hash_key); 3887 3888 #endif 3889 3890 void tcp_done(struct sock *sk) 3891 { 3892 struct request_sock *req; 3893 3894 /* We might be called with a new socket, after 3895 * inet_csk_prepare_forced_close() has been called 3896 * so we can not use lockdep_sock_is_held(sk) 3897 */ 3898 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1); 3899 3900 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 3901 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 3902 3903 tcp_set_state(sk, TCP_CLOSE); 3904 tcp_clear_xmit_timers(sk); 3905 if (req) 3906 reqsk_fastopen_remove(sk, req, false); 3907 3908 sk->sk_shutdown = SHUTDOWN_MASK; 3909 3910 if (!sock_flag(sk, SOCK_DEAD)) 3911 sk->sk_state_change(sk); 3912 else 3913 inet_csk_destroy_sock(sk); 3914 } 3915 EXPORT_SYMBOL_GPL(tcp_done); 3916 3917 int tcp_abort(struct sock *sk, int err) 3918 { 3919 if (!sk_fullsock(sk)) { 3920 if (sk->sk_state == TCP_NEW_SYN_RECV) { 3921 struct request_sock *req = inet_reqsk(sk); 3922 3923 local_bh_disable(); 3924 inet_csk_reqsk_queue_drop(req->rsk_listener, req); 3925 local_bh_enable(); 3926 return 0; 3927 } 3928 return -EOPNOTSUPP; 3929 } 3930 3931 /* Don't race with userspace socket closes such as tcp_close. */ 3932 lock_sock(sk); 3933 3934 if (sk->sk_state == TCP_LISTEN) { 3935 tcp_set_state(sk, TCP_CLOSE); 3936 inet_csk_listen_stop(sk); 3937 } 3938 3939 /* Don't race with BH socket closes such as inet_csk_listen_stop. */ 3940 local_bh_disable(); 3941 bh_lock_sock(sk); 3942 3943 if (!sock_flag(sk, SOCK_DEAD)) { 3944 sk->sk_err = err; 3945 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 3946 smp_wmb(); 3947 sk->sk_error_report(sk); 3948 if (tcp_need_reset(sk->sk_state)) 3949 tcp_send_active_reset(sk, GFP_ATOMIC); 3950 tcp_done(sk); 3951 } 3952 3953 bh_unlock_sock(sk); 3954 local_bh_enable(); 3955 tcp_write_queue_purge(sk); 3956 release_sock(sk); 3957 return 0; 3958 } 3959 EXPORT_SYMBOL_GPL(tcp_abort); 3960 3961 extern struct tcp_congestion_ops tcp_reno; 3962 3963 static __initdata unsigned long thash_entries; 3964 static int __init set_thash_entries(char *str) 3965 { 3966 ssize_t ret; 3967 3968 if (!str) 3969 return 0; 3970 3971 ret = kstrtoul(str, 0, &thash_entries); 3972 if (ret) 3973 return 0; 3974 3975 return 1; 3976 } 3977 __setup("thash_entries=", set_thash_entries); 3978 3979 static void __init tcp_init_mem(void) 3980 { 3981 unsigned long limit = nr_free_buffer_pages() / 16; 3982 3983 limit = max(limit, 128UL); 3984 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */ 3985 sysctl_tcp_mem[1] = limit; /* 6.25 % */ 3986 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */ 3987 } 3988 3989 void __init tcp_init(void) 3990 { 3991 int max_rshare, max_wshare, cnt; 3992 unsigned long limit; 3993 unsigned int i; 3994 3995 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE); 3996 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > 3997 sizeof_field(struct sk_buff, cb)); 3998 3999 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL); 4000 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL); 4001 inet_hashinfo_init(&tcp_hashinfo); 4002 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash", 4003 thash_entries, 21, /* one slot per 2 MB*/ 4004 0, 64 * 1024); 4005 tcp_hashinfo.bind_bucket_cachep = 4006 kmem_cache_create("tcp_bind_bucket", 4007 sizeof(struct inet_bind_bucket), 0, 4008 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 4009 4010 /* Size and allocate the main established and bind bucket 4011 * hash tables. 4012 * 4013 * The methodology is similar to that of the buffer cache. 4014 */ 4015 tcp_hashinfo.ehash = 4016 alloc_large_system_hash("TCP established", 4017 sizeof(struct inet_ehash_bucket), 4018 thash_entries, 4019 17, /* one slot per 128 KB of memory */ 4020 0, 4021 NULL, 4022 &tcp_hashinfo.ehash_mask, 4023 0, 4024 thash_entries ? 0 : 512 * 1024); 4025 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) 4026 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 4027 4028 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 4029 panic("TCP: failed to alloc ehash_locks"); 4030 tcp_hashinfo.bhash = 4031 alloc_large_system_hash("TCP bind", 4032 sizeof(struct inet_bind_hashbucket), 4033 tcp_hashinfo.ehash_mask + 1, 4034 17, /* one slot per 128 KB of memory */ 4035 0, 4036 &tcp_hashinfo.bhash_size, 4037 NULL, 4038 0, 4039 64 * 1024); 4040 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 4041 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 4042 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 4043 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 4044 } 4045 4046 4047 cnt = tcp_hashinfo.ehash_mask + 1; 4048 sysctl_tcp_max_orphans = cnt / 2; 4049 4050 tcp_init_mem(); 4051 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 4052 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 4053 max_wshare = min(4UL*1024*1024, limit); 4054 max_rshare = min(6UL*1024*1024, limit); 4055 4056 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM; 4057 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024; 4058 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare); 4059 4060 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM; 4061 init_net.ipv4.sysctl_tcp_rmem[1] = 131072; 4062 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare); 4063 4064 pr_info("Hash tables configured (established %u bind %u)\n", 4065 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 4066 4067 tcp_v4_init(); 4068 tcp_metrics_init(); 4069 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0); 4070 tcp_tasklet_init(); 4071 mptcp_init(); 4072 } 4073